Audio driven facial animation using machine learning

Abstract

The invention discloses audio-driven facial animation using machine learning. Systems and methods of the present disclosure include animating an avatar or proxy according to input audio and one or more selected or determined emotions and / or styles. For example, a deep neural network may be trained to output motion or deformation information of a character representing that the character spokes an utterance contained in the audio input. The character may have different facial components or regions (e.g., head, skin, eyes, tongue) that are modeled individually, such that the network may output motion or deformation information for each of these different facial components. During training, the network may train associated decoders using weighted feature vectors using converter-based audio encoders with locked parameters. The network output may be provided to a renderer to generate a mood-accurate audio-driven facial animation.

Description

Background Art

[0001] It may be desirable for each operation to animate a character to appear as if the character is speaking the speech represented by the audio data. Partly due to the time and complexity of creating such animations, it may be beneficial to automate such a process, particularly for real-time or near real-time operations. Machine learning-based methods have been used to generate animations of characters based on input audio, but these existing methods are generally limited in their capabilities, resulting in animations that are not realistic enough in many instances. For example, existing methods may attempt to animate various facial features of a character, including a mouth or eyes, so as to correspond to the speech represented by the corresponding audio data, but when used on a language for which a model is not explicitly trained, these models generally cannot provide realistic animations. For operations in which a character is a virtual person (intended to make the virtual person appear as if it were an actual person speaking a speech in a realistic manner with realistic behavior), this problem may be exacerbated. BRIEF DESCRIPTION OF THE DRAWINGS

[0002] Various embodiments according to the present disclosure will be described with reference to the accompanying drawings, in which:

[0003] Figure 1A and Figure 1B Various aspects of a character that may be animated based at least in part on speech data are shown in accordance with at least one embodiment;

[0004] Figure 2 An example network for generating animations corresponding to speech (including emotional and style support) is shown in accordance with at least one embodiment;

[0005] Figure 3 An example network for generating an animation corresponding to an utterance is shown in accordance with at least one embodiment;

[0006] Figure 4 An example training process for a network to generate animations of a character corresponding to input audio is shown in accordance with at least one embodiment;

[0007] Figure 5A and Figure 5B shows an example reasoning process for a network to generate an animation of a character corresponding to input audio in accordance with at least one embodiment;

[0008] Figure 6 Components of a distributed system that may be utilized to update or perform inference using a machine learning model in accordance with at least one embodiment are shown;

[0009] Fig. 7A Inference and / or training logic according to at least one embodiment is shown;

[0010] Figure 7B Inference and / or training logic according to at least one embodiment is shown;

[0011] Figure 8 An example data center system is shown in accordance with at least one embodiment;

[0012] Fig. 9 A computer system according to at least one embodiment is shown;

[0013] Fig.10 A computer system according to at least one embodiment is shown;

[0014] Fig.11 illustrates at least a portion of a graphics processor according to one or more embodiments;

[0015] Fig.12 illustrates at least a portion of a graphics processor according to one or more embodiments;

[0016] Fig.13 is an example data flow diagram of a high-level computing pipeline according to at least one embodiment;

[0017] Fig.14 is a system diagram of an example system for training, adapting, instantiating, and deploying machine learning models in an advanced computing pipeline according to at least one embodiment; and

[0018] Fig.15A and Fig. 15B A data flow diagram of a process for training a machine learning model, and a client-server architecture for augmenting an annotation tool with a pre-trained annotation model, according to at least one embodiment are shown. DETAILED DESCRIPTION

[0019] In the following description, various embodiments will be described. For the purpose of explanation, specific configurations and details are set forth in order to provide a thorough understanding of the embodiments. However, it will also be apparent to those skilled in the art that these embodiments can be practiced without these specific details. In addition, well-known features may be omitted or simplified so as not to obscure the described embodiments.

[0020] The systems and methods described herein may be used by, but are not limited to, non-autonomous vehicles or machines, semi-autonomous vehicles or machines (e.g., in in-cabin infotainment or digital or driver virtual assistant applications), autonomous vehicles or machines, driverless and driverless robots or robotic platforms, warehouse vehicles, off-road vehicles, vehicles coupled to one or more trailers, airships, boats, shuttles, emergency response vehicles, motorcycles, electric or motorized bicycles, aircraft, construction vehicles, trains, submarines, remotely controlled vehicles (such as drones), and / or other vehicle types. Further, the systems and methods described herein may be used for a variety of purposes, including, by way of example and not limitation, for machine control, machine motion, machine driving, synthetic data generation, model training or updating, perception, augmented reality, virtual reality, mixed reality, robotics, safety and surveillance, simulation and digital twins, autonomous or semi-autonomous machine applications, deep learning, environmental simulation, object or actor simulation and / or digital twins, data center processing, conversational artificial intelligence (AI), generative AI with a large language model (LLM), light transport simulation (e.g., ray tracing, path tracing, etc.), collaborative content creation for 3D assets, cloud computing, and / or any other suitable application.

[0021] The disclosed embodiments may be included in a variety of different systems, such as automotive systems (e.g., control systems for autonomous or semi-autonomous machines, perception systems for autonomous or semi-autonomous machines), systems implemented using robots, aerial systems, medical systems, rowing systems, smart area monitoring systems, systems for performing deep learning operations, systems for performing simulation operations, systems for performing digital twin operations, systems implemented using edge devices, systems including one or more virtual machines (VMs), systems for performing synthetic data generation operations, systems implemented at least in part in a data center, systems for performing conversational AI operations, systems for performing generative AI operations using LLMs, systems for performing light transport simulations, systems for performing collaborative content creation of 3D assets, systems implemented at least in part using cloud computing resources, and / or other types of systems.

[0022] Methods according to various embodiments may generate animations representing one or more characters speaking the speech represented by the audio data. This may include, for example, high-resolution, full three-dimensional (3D) facial animations to be presented as part of a movie, game, artificial intelligence (AI)-based agent, digital avatar, teleconferencing, virtual / augmented / mixed / enhanced reality experience, and / or other media presentation, content, or experience. One or more deep neural networks (such as frame-based convolutional neural networks (CNNs) or recurrent neural networks (RNNs) and / or transformer-based models) may take raw audio as input, extract features from the raw audio, and receive one or more component vectors, with which the character will be animated to speak the speech contained in the audio segment extracted from the input raw audio. The network may then provide outputs, such as motion, vertex, and / or deformation data, which may be provided to a renderer, for example, to generate or synthesize facial animations corresponding to that portion of the speech. During training, in addition to the audio input, the network may receive one or more component vectors, such as a style vector or an emotion vector (which indicates one or more emotions with potential relative weights), for rendering facial animations for the input audio clip. The one or more emotions indicated by the data may change at various points or keyframes in the audio data. The network may also receive a style vector (or style information contained in an emotion vector or other emotional data representation) indicating modifications or fine control of the animation to be generated for the indicated one or more emotions, such as may relate to the animation style or may relate to specific movements to be modified or enhanced, as well as other such options. The motion or deformation information output by the network may correspond to a set of facial (or other body) components or parts that can be animated independently at least to some extent to realistically represent the character speaking the input speech. These components may include, for example, the head, jaw, eyeball, tongue, or skin of the character. In an embodiment, in addition to or in place of facial components or parts, body components or parts (such as arms, legs, torso, neck, etc.) may be modeled. Modeling each of these facial (and / or body) components individually, and determining deformations for each of these components, can make the rendered facial (and / or body) animation appear more realistic for a given emotion, particularly when taking into account any style data provided to the network.

[0023] Various systems and methods are directed to incorporating one or more converter-based audio encoders into the Audio2Face model. Various embodiments make changes to the formant analysis network of the Audio2Face model to modify an existing CNN-based system with a pre-trained converter model. In at least one embodiment, the modification includes, at least in part, incorporating one or more converter layers into the audio encoder model, which may also include one or more CNN layers. The model may be pre-trained on audio data that may contain a variety of different languages, and then the model may be used within the Audio2Face model to help train a decoder to generate a vector set for use by one or more application-specific outputs such as within a renderer to generate an animated face (e.g., a character) that reproduces how a human face can say an input audio sample. The pre-trained converter model may include multiple layers, and given a set of input features from a CNN layer, each layer may generate a set of vectors associated with those features. The system and method add a combination layer, which may include a learnable weight, which may be applied to linearly combine each of the output vectors of each layer to generate the final output. When the model is trained by back propagation, the weight of the combination layer can be learned. Various embodiments can be used with audio samples of small duration (e.g., approximately less than 1 second, approximately less than 0.8 seconds, approximately less than 0.6 seconds, approximately between 0.4 seconds and 0.6 seconds, approximately between 0.6 seconds and 0.8 seconds, approximately between 0.8 seconds and 1 second, or any other reasonable duration), and information from previous frames may not be incorporated. Therefore, the user can select any part of the larger audio clip for evaluation and viewing without losing output quality due to lack of knowledge of the previous frame. In addition, techniques for improving output quality can be combined. In at least one embodiment, additional regularization losses are incorporated to reduce motion jitter during silence. For example, if the sound volume is low, this loss can penalize large motion between adjacent animation frames. In addition, one or more embodiments may also incorporate a lip distance loss to select key points on the upper / lower lip of the character and then minimize the difference between the predicted distance and the true distance between the upper / lower lip key points.

[0024] When generating such image or video data for various operations, it may be a goal (or in some examples a requirement) for the representation of a character (such as a person, robot, animal, or other such entity) to behave as realistically as possible. Such realistic behavior may include various movements or actions in various states and under various conditions. For example, a character (such as a Figure 1A100) can be animated to move their mouth, face and / or head in a manner that conveys that the character is speaking the speech represented by audio data, and the audio data can be provided to be played or otherwise presented with the animation. In order to make the animation look as realistic as possible, it may not be enough to just move the character's mouth to appear to speak the words in the speech, because the set of emotional states or environments may affect how the character physically moves other elements of its face (e.g., eyes, tongue, etc.) when speaking the speech. For example, the character can be animated to speak the words "I'm here". Simply animating the mouth to correspond to the formation of the words is not enough to convey authenticity in various situations, because the user may speak those words with very different intentions, emotions and / or styles. For example, as shown in the first image 102, the character can be animated to speak words with a happy emotion, such as when the character has returned to their family and is happy to see them. In the second image 104, the character may be angry that they are called to a certain location, and may express anger when conveying that they have arrived. In the third image 106, the character may feel a sense of disgust, such as in the case where the character is in a place they really don't want to visit but are forced to visit. In the fourth image 108, the character may feel a sense of sadness because the character may announce their arrival, but have felt the need to arrive due to unexpected circumstances. There may be any of many other emotions that the character may experience that are to be conveyed in different situations. In order to make the animation look realistic, the animation system (or other image data generation or synthesis system, component, module, or device) can accept or infer emotional states and attempt to generate an animation for the character that not only matches any audio to be spoken by the character, but also conveys speech with emotional behavior.

[0025] As shown, the communication of emotional behavior can include multiple different but related movements. For example, the outer surface of the user's head (corresponding to Figure 1B The skin 152 on the head shown in the figure can be deformed to convey words and emotions. This can include moving the lips 154 of the character to match the formation of the words being spoken so that the animation of the character appears to be speaking the words. As mentioned, the lip movement can also be determined based at least in part on the emotional state of the character when the words are spoken. There can also be other aspects of the skin (or head) that can also exhibit related behaviors based at least in part on emotional states. For example, the amount or type of head movement (e.g., rotation or tilt of the head) can vary with different emotional states. Further, aspects such as certain wrinkles 156 or lines in the skin of the character can change for different emotions, such as becoming more prominent or less prominent or taking on a specific shape. Similar emotion-based behaviors can also be exhibited by other skin areas or features (such as by eyebrows 158, cheeks, etc.).

[0026] In addition to this outer skin or surface, there may also be other aspects or features of this character that can change behavior with different emotions, which may only be somewhat related to the behavior of the skin or surface, as may be caused by the physical or kinematic constraints of the character. For example, the eyes 160 of the character can be modeled separately from the face at least to some extent. The position 162 of the eyes can depend on the position of the head or skin of the character, but because the position of the eyes is relatively fixed in the eye sockets of the character, the movement or orientation of the eyes can be independent of the behavior of the skin at least to some extent. For example, if angry, the character may focus directly on the person they are talking to, and if feeling sad or guilty, the character may not look at another person. Similarly, the amount of panning movement or the frequency at which the character changes the focus of its eyes can be changed for different emotions. Therefore, it can be expected to reason eye orientation separately from skin, head or surface behavior at least to some extent.

[0027] Conventional eye tracking solutions may not provide adequate performance. In at least one embodiment, pupil tracking can be performed from input 3D capture data using an algorithm such as the Lucas-Kanade optical flow algorithm, which provides a differential method for optical flow estimation, assuming that the optical flow is essentially locally constant, and solves for the basic optical flow within the local neighborhood. In the event of a blink or an obstruction, or in the event that at least one eye is no longer visible in the captured image data, at least some amount of interpolation can be performed based on one or more preceding (or subsequent, if available) image frames. Such eye tracking methods can also capture small saccadic movements of the eyes, which can help make eye movements appear more natural in rendered facial animations. Such methods can accurately model eye movements without the need for image data representing images that focus primarily on the eyes of the actor when the actor speaks.

[0028] Similarly, the character's tongue 164 can move independently of the head to at least some extent within physical or kinematic constraints. The amount or type of tongue movement can vary with emotion, as a sad character may exhibit very little tongue movement, while an angry or excited character may exhibit a lot of tongue movement, which may also differ in direction or pattern. In addition, the movement of the tongue 164 may also be related to the language spoken by the character, as different sounds may not exist in every language. A suitable number of feature points may be used for the tongue mesh, allowing realistic movement and behavior to be achieved through, for example, mesh deformation. The number of points may be reduced or compressed (to a number such as, but not limited to, 10 points) through a process such as principal component analysis (PCA) to reduce the amount of processing and memory required for tongue mesh deformation.

[0029] There may also be other aspects or features of the character, which may also be modeled separately to improve realism. For example, the jaw 166 of the character may be modeled separately from the head of the character. Although the movement of the jaw may be able to be approximated by skin movement and deformation, it is observed that for at least some systems or implementations, this reasoning may not be accurate enough to avoid any post-processing or manual cleanup of the generated animation. In order to improve accuracy, the movement of the jaw 166 may be modeled separately, because the jaw can move different amounts in many different directions for similar states of the skin, such as when the lips of the character are closed, but it may be difficult to capture the movement based only on skin deformation. There may be other aspects, features or components of the character that may also benefit from being modeled separately, which may depend at least in part on the type of the character, because animals, robots or aliens may be modeled as having different skeletal structures or kinematic capabilities. Different types or instances of the same character may also exhibit different behaviors or different emotions, such as people of different ages, genders, backgrounds, or other such aspects.

[0030] In many cases, the user may not only show a single emotion, or may show one or more emotions of different levels. For example, for the "angry" emotion type, if the character is slightly uneasy rather than the character is irritated, the character may behave very differently. The character can also show multiple emotions at once, such as the character who is happy that the child is admitted to the university and the sad child will leave, and therefore will realistically show the characteristics associated with the combination of these two emotions. At least in some cases, there may also be a character with different styles of behavior for the same emotion. For example, if you talk to a stranger instead of talking to a partner, parent or child, the character can act differently. If in a professional setting, the character may also act differently from in a personal setting. In some instances, such as for a game or a movie, an animator may simply want the character to show a specific appearance, style or behavior for a certain emotional state. Therefore, for at least some methods presented herein, it is beneficial to allow users (or applications or operations, etc.) to specify more than one emotion or a combination of emotions. In some operations, users (or other sources) can also specify the weighting of these individual emotions to provide a more accurate combination of emotions. The user may also be able to specify different emotions, combinations, or weights at different points in time, or emotional "keyframes" in an animation, such as where a character may become increasingly sad or may calm down during a discussion. The user may also be able to specify the style in which the character conveys the emotions, which may also change over time, such as at different keyframes in an animation.

[0031] Methods according to various embodiments may use at least some of these and other such aspects or features to provide facial animation that provides realistic behavior for various different character types and for various different input audio types in various emotional states. This may include, for example, full three-dimensional (3D) facial animation driven by audio with emotional control. In this method, realistic animations can be generated without any manual input or post-processing required - although it is possible in the desired case. Automating such animations can help significantly reduce the amount of time, experience and cost required for manual (or at least partially manual) character animations. Compared to traditional methods, audio-driven facial animation can provide an efficient way to generate facial animations because only audio data is required to drive the animation of a given character. Previous attempts at audio-driven animation can animate the lower face to synchronize with the lips, but cannot generate the appropriate movement or behavior of other facial areas or features (such as the upper face, teeth, tongue, eyes and head) that may be required for accurate behavior representation. In existing methods, it is usually necessary to use additional manual or post-processing work to correct inaccurate behavior in the generated animation. Prior attempts to include emotions in animations for speech have also typically focused only on a single type of emotion for the duration of the speech, which in many cases does not capture or accurately represent the natural variations or shifts in behavior, which then also typically requires additional manual or post-processing work. Furthermore, the models used in existing approaches often fail to generalize across different languages ​​or audio inputs, such as non-verbal inputs (e.g., yelling, gasping, sobbing, etc.). In attempts to address these issues, existing approaches have incorporated additional training data for language-specific applications or added application-specific loss functions, which increases the complexity and cost of facial animation.

[0032] The method according to various embodiments can provide automatic audio-driven animation (such as full 3D facial animation) with variable emotion control, which can be generalized on different languages ​​and / or non-language audio inputs. In at least one embodiment, one or more aspects of the animation pipeline are configured to include a pre-trained audio encoder, which includes one or more converter layers. The encoder can be trained on various different languages ​​and / or multilingual audio clips (slips) to calculate audio features from input audio clips on the timeline. The calculated audio features from each converter layer can be linearly combined using learnable weights. The combined audio features can then be passed along the pipeline to one or more audio decoders for estimating various animation coefficients, which can be used by one or more renderers to render characters using facial movements associated with audio input. In at least one embodiment, the pre-trained audio encoder is frozen during training, which can reduce the possibility of overfitting, and therefore during training, only the decoder and linear combination layer weights can be modified. The system and method can be used offline and in real time and / or near real time (e.g., without significant delay).

[0033] In at least one embodiment, a collection of speech performances of one or more actors speaking speech (e.g., specific sentences) in different languages, different non-verbal audio, emotions, emotion levels, emotion combinations, or presentation styles, as well as other such options, may also be captured. The emotions supported by such a system (as an example of an input component vector) may include any appropriate emotion (or similar behavior or state) that can be represented at least in part by character animation, image synthesis, or rendering, such as joy, anger, surprise, sadness, pain, or fear, etc. The data collection process may include capturing, for example, 4D data that includes multiple views over a time period of at least the speech being spoken. Figure 3 D data. Reconstruction of such captured facial behavior may be performed not only for facial skin (or such surface) but also for other articulatory or controllable components, elements or features (e.g., which may include teeth, eyeballs, head and tongue (and / or body features or components, such as limbs, fingers, toes, torso, etc.)). The reconstruction may provide geometric deformation data in the time domain for each separately (or at least somewhat separately) modeled facial (or other body) component or region. Such reconstruction may provide a complete data set for training, for example, a deep neural network 206 (e.g., Figure 2 ), to perform tasks such as 3D facial animation.

[0034] In at least one embodiment, the frame-to-frame mapping can be used to generate a single animation frame output. For example, a single frame of audio (e.g., approximately 0.52 seconds) can be processed with a corresponding single animation frame output. Figure 2In the example system 200 of , an audio input 202, which may include one or more of the original audio, audio frames, audio clips within the current audio window, etc., may be provided as an input to a deep neural network, which may use an analysis network portion 208 to analyze the audio and encode features representing the features of the audio in the audio data 202 (such as a portion of the speech). The analysis network portion 208 may include one or more audio decoders and one or more audio encoders for encoding the audio features into a feature vector, which, as indicated herein, may be one or more combined audio features from the converter layer. The feature vector may be provided as an input to the pronunciation network portion 210 of the deep neural network 206. In this example, a component vector 204 may be provided as an input. In some embodiments, multiple vectors may be provided, such as a "mood" vector and a "style" vector, as well as various other options and combinations. In addition, such as by using a fusion process, a single vector may be provided as an input including both style and / or emotion data. The emotion vector may include data of one or more emotions applied to the speech being used for training, such as indicating the emotions used by the voice actor when speaking the speech captured in the audio data. In some cases, this may include data for a single emotion label, such as "angry," or may include data for multiple emotions, such as "angry" and "sad," and potential relative weights of those two emotions. These labels and / or weights may have been provided to the voice actor initially, may have been determined after the utterance was spoken, and / or may include updated labels after hearing the utterance spoken for a particular emotion in an audio capture, among other such options.

[0035] In at least one embodiment, the component vector 204 may correspond to a style vector that is provided as an input to the deep neural network 206 during training (and similarly in deployment). The style vector may contain data about any aspect of the animation or motion of a facial component that modifies how one or more points of one or more facial components should move for a given emotion or emotion vector. This may include affecting the motion of a particular feature or facial component, or providing a style of the overall animation to be used, such as "tense" or "professional". The style vector may also be viewed as a finer-grained control over the emotion, where the emotion vector provides a label for the emotion to be used, and the style provides a finer-grained control over how the emotion is expressed through the animation. Other methods of determining style data may also be used. In different implementations, a single set of emotion and style vectors may be provided for a given audio clip, a set of vectors may be provided for each frame of an animation to be generated, or a set of vectors may be provided for a particular point or frame of an animation (e.g., an emotion keyframe) where at least one emotion or style value or setting is to be modified relative to a previous frame.

[0036] In this system 200, the component vectors 204 are fed into a pronunciation network portion 210 of a deep neural network 206 at multiple levels, including at least the beginning and end of the network, to help adjust the network. The deep neural network 206 can use a shared audio encoder and one or more (e.g., multiple) decoders for each facial component (e.g., facial skin, jaw, tongue, eyeball, and head). During training, the output network portion 212 of the deep neural network 206 can generate a set of head / jaw displacements, eyeball rotations, and skin / tongue vertex positions / or PCA coefficients of vertices 214 and / or motion vectors (or other motions or deformations) for each feature point of the facial component, whether for each such feature point or only for those feature points that have changed relative to the previous frame, and other such options. During training, these vertex positions / or PCA coefficients can be compared with "true value" data (such as original reconstructed facial data from (e.g., 4D) image capture) to calculate a total loss value. In at least one embodiment, a loss (such as an L2 loss) can be used for both the position and velocity of the feature points in the output data representation. In at least some embodiments, the loss function used to determine the loss value may include terms for position, motion, and adversarial losses. This loss value may be used during backpropagation to update the network parameters (e.g., weights and biases) of the deep neural network 206. As described herein, in at least one embodiment, the weights of the audio encoder may be frozen so that backpropagation is only used to update one or more decoders. Once it is determined that the network has converged to an acceptable or desired level of accuracy or precision and / or another end-of-training criterion (e.g., processing all training data or performing a target number or maximum number of training iterations) is met, the trained network 206 may be provided or deployed for inference.

[0037] During inference, the network can receive audio input 202 (e.g., in some embodiments only audio data) as input, and can infer a set of vertex positions 214 for each facial component (e.g., head, face, eyeballs, jaw, tongue), which can then be fed to a renderer 216 (e.g., a rendering engine of an animation or video synthesis system) to generate an animation frame 218, which can be one of a series of frames that provide an animation when rendered or played. As discussed in more detail elsewhere herein, if the generated vertex positions are to be modified in some way with respect to how the deep neural network 206 would otherwise infer vertex positions based on audio data, then emotion or style vector data can also be provided as input to the deep neural network 206, such as to convey a particular style or facial behavior to be used in inferring the vertex positions 214.

[0038] As described in more detail later herein, the deep neural network 206 can output a vector encoding the position or motion data of each point on a mesh of one or more facial components, and this output vector (or another output, such as a global transformation matrix) can be fed to a renderer 216, which can apply these values ​​to one or more meshes of this character to guide the animation. The output matrix or vector can have dimensions that match the characteristics of the facial components, such as can include, for example but not limited to, 272 facial feature points for skin, 5 facial feature points for head, 5 facial feature points for jaw, 2 facial feature points for eyeballs, 10 facial feature points for tongue (e.g., using PCA compression), etc. Such an approach can provide sufficiently smooth animations that, in at least many cases, no additional smoothing or post-processing will be required. However, the system can allow additional smoothing to be applied, such as where the user can specify one or more smoothing parameters.

[0039] The size of the audio data and / or the window of audio data can be any appropriate or suitable size and can depend in part on the implementation, but can include at least a time period corresponding to an animation frame for a target frame rate (e.g., 60 Hz), and can include larger windows to consider portions of the audio for nearby frames (e.g., before or after) to provide more accurate and smooth animation and more accurate determination of emotion and / or style from the input audio. The example system can use one-hot vector encoding to represent different emotions or emotion labels, where the resulting emotion vectors can be concatenated at one or more layers.

[0040] In at least one embodiment, the output of such a network may include data specifying motion for different facial parts or components. The output may also include data for components of the character other than facial components, such as arms, legs, torsos, etc. This system may also generate accurate motion data for the character for frames or scenes where the character's face is not visible or is only partially represented in the scene. For facial parts with non-rigid deformations (such as skin and tongue), principal component analysis (PCA), for example, may be used to compress the captured (e.g., 4D) motion data. This may allow a facial mesh with a large number of points (such as 60,000 points) to be represented by a vector of much smaller dimensions, such as may correspond to 272 (or another number of) eigenvalues. In such embodiments, the PCA weight vector may be used as a training representation. In one embodiment, a fully connected layer may fuse emotion and / or style data into a smaller vector, which may be inserted into each layer of the network as a concatenation.

[0041] For facial parts with rigid transformations (such as the head and teeth), multiple (for example, but not limited to) Figure 1B 5) markers or feature points shown in , and the position increments of these points to a reference position can be used as a training representation, from which a rigid transformation matrix can be calculated. For the mandible, the five feature points may include a point at either end of the mandible, a center point, and two intermediate reference points, where those intermediate points may not be necessary, but may help with fine motion control and noise reduction. For rotatable components such as eyeballs, the system can use two rotation values ​​(e.g., pitch and yaw) to represent horizontal and vertical rotations relative to a default orientation. Full 3D facial animation output can be obtained by such a system at runtime or during inference. Such a system can also allow the emotion or style of the speech animation to be interactively controlled by feeding different component vectors 204 into the network to allow modification or "fine-tuning" of the motion of each of the individual facial components and facial regions, such as can be part of a real-time or near real-time process, or as post-processing.

[0042] In at least one embodiment, the deep neural network 206 can have an architecture based on a per-frame convolutional neural network (CNN). The per-frame CNN architecture can receive, for example but not limited to, an audio window of approximately 0.5 seconds as input, which can include data from previous and subsequent frames in a sequence, from which the CNN can predict data for intermediate frames in the sequence. Other architectures can also be used within the scope of different embodiments, which can also provide smoothness of animation based at least in part on context provided to or determined by those architectures. Although different architectures can provide adequate results, some architectures can perform better in certain situations or for certain aspects of speech-driven facial animation. For example, CNN-based architectures perform well for real-time reasoning and can generate very reliable lip sync movements.

[0043] In at least one embodiment, the deep neural network can be a converter architecture. The converter architecture can weight each part of the input differently. The input data can be provided as a sequence, but can be processed all together, where the attention mechanism provides context for the position of the input sequence. The converter may include an encoder architecture, where the encoding layer iteratively processes the input. The encoder determines the relevant features of the input and passes these encodings to the next layer. The attention mechanism can be incorporated to weight the relevance of each part of the input and / or output. For example, the attention mechanism can accept the input encodings from the previous encoder and weight their relevance to each other to generate the output encodings. The encoder can then further process each output encoding separately in subsequent layers. These output encodings are then passed to the next encoder as its input.

[0044] Figure 3An example system 300 is shown representing one or more components forming an analysis network 208, which may be provided in at least one embodiment. In this example, an audio input 202 is received at a deep neural network 206. In at least one embodiment, the audio input 202 is a raw audio input, however, various embodiments may include an audio input that has been subjected to one or more pre-processing or post-processing operations. In this example, the audio input 202 is provided to a CNN 302, which may form a part of the analysis network 208. The CNN 302 may be part of or at least provide information to a multi-layer convolutional feature encoder that receives the audio input 202 and outputs one or more potential representations 304. The potential representation 304 may be provided for one or more time steps and fed to a converter 306 to determine a representation of an information sequence associated with the potential representation 304 and generate a series of context representations 310. In at least one embodiment, each layer 308 of the converter 306 may generate a set of context representations 310 for the potential representation 304 provided by the CNN 302. For example, various embodiments may include approximately 25 different features, wherein a contextual representation would be developed for each of the 25 features for each layer 308 .

[0045] In at least one embodiment, the combination layer 312 may include learnable weights that may be applied to linearly combine each of the output vectors of each layer 308 to generate an output 314. When the model is trained via back-propagation, the weights of the combination layer 312 may be learned. This output 314 may then be used by the pronunciation network 210, which may also receive Figure 2 One or more component vectors 204 are shown to generate position information for rendering a character model.

[0046] Various embodiments mask these latent representations 304 before feeding them into the transformer 306. From there, contextualized representations can be built, such as via a contrastive learning process that compares different samples to each other. The model can be pre-trained on unlabeled data and then fine-tuned with labeled data, among other options. Thus, the systems and methods can use one or more models to build contextual representations using self-attention to capture dependencies.

[0047] In at least one embodiment, CNN 302 includes time convolution, followed by layer normalization and activation functions. Audio input 202 can be a raw waveform that is normalized before processing. Converter 306 can receive output feature encoding. Each layer 308 can learn multiple features and output vectors for those associated features. The total number of features of each layer 308 can be about 10 features, about 15 features, about 20 features, about 25 features, about 30 features, or any reasonable number of features. In addition, the number of features can be in the range of about 10 to 20 features, about 15 to 25 features, about 20 to 30 features, or any other reasonable range.

[0048] At the combination layer 312, each vector of each layer 308 can be weighted and then linearly combined to generate an output 314. The weights can be learned during training and / or can be adjusted or tuned based on different applications. The weighted vectors indicate which features are more useful to the converter layer 308, and the features are linearly combined based on this determination. For example, if the feature from layer 1 is f1 and the feature from layer 2 is f2, these features are fed into the combination layer 312, and the output is calculated as f = (f1*w1+f2*w2) / (w1+w2), where w1 and w2 are learnable weights.

[0049] In at least one embodiment, the audio input 202 can be a short audio sequence, such as an audio sequence that is less than 1 second long. For example, the audio input 202 can be about 0.5 seconds long. Such a small audio sample can provide benefits for implementing real-time or near real-time applications. In addition, a given frame or audio clip can be unrelated to or not otherwise dependent on a previous frame or audio clip. Therefore, the output animation can be "skipped" or tracked along different parts of the audio clip without losing information that would be needed if it were dependent on a previous frame.

[0050] Various embodiments of the present disclosure may also implement one or more techniques to improve output quality. For example, an additional regularization loss may be added to reduce motion jitter during silence. If the sound volume is low, the loss will penalize large motions between adjacent animation frames. Thus, there may be a reduced likelihood of movement between frames, which would be desirable if there were no sound, since the characters may not move their mouths, for example, if they were not speaking. Similarly, the systems and methods may also apply a lip distance loss to improve output quality. For example, key points may be selected on the lips (e.g., upper and lower lips), and then the minimum difference between the predicted distance and the true distance between the key points may be selected. When the true lip distance is small (e.g., when the mouth is closed), the loss may have a higher weight. As a result, the output character model may provide improved realism by guiding or otherwise encouraging the mouth to close at the appropriate time.

[0051] One or more embodiments of the present disclosure may train the converter 306 so that the parameters of the encoder are locked during training, but the parameters of the associated decoder (e.g., a convolutional decoder) are allowed to change as a result of the training. For example, a pre-trained model trained on different audio data (e.g., a variety of different languages ​​or sounds) may be used, and the associated parameters of the encoder may not change or otherwise be updated during the training process. In this way, overfitting may be reduced and a smaller amount of training data may be used.

[0052] Embodiments of the present disclosure incorporate the illustrated converter 306 into the deep neural network 206 (such as, into the analysis network 208) for processing one or more audio inputs 202, which may correspond to windows or clips of audio segments. As noted herein, the audio segments may be original audio segments or may be processed audio segments. In addition, in at least one embodiment, the audio input may be cached or otherwise clipped before being processed within the deep neural network 206 and / or as part of the deep neural network 206 in one or more preprocessing stages. As an example, the deep neural network 206 may use an audio buffer of a given length, such as length 8320 (e.g., approximately 0.52s of audio), to calculate audio features along the time axis. In some embodiments, the number of audio features is determined by one or more properties of the CNN 302 and / or the converter 306. For example, existing systems may evaluate approximately 50 different features. Various embodiments of the present disclosure may evaluate more or fewer features, such as approximately 25 features. These calculated features may then be evaluated at each layer in the layer 308 of the converter 306. Thereafter, the calculated audio features from each converter layer 308 may be linearly combined using learnable weights (e.g., at the combination layer 312), wherein the weights may be learned during one or more training operations. In addition, in various embodiments, the weights may be adjusted for a particular application based on one or more desired outputs. The linearly combined audio features are then passed to an audio decoder for estimating animation coefficients. As described above, in at least one embodiment, although the parameters of the audio decoder may be adjusted or modified based on training information, the system and method may use a pre-trained audio encoder that is locked or otherwise unmodifiable during training. That is, the pre-trained audio encoder is frozen and only the decoder weights are trained. In at least one embodiment, the decoder may be a convolutional decoder. Therefore, due to the non-autoregressive nature, the system and method may be used for both online and streaming (e.g., in real time and / or near real time).

[0053] The systems and methods of the present disclosure may be incorporated into one or more user interfaces that may be used to adjust or otherwise modify the facial or body animation of one or more computer-animated characters. For example, a user interface may be provided that shows an animated representation of a character at different times, and one or more inputs may be provided to the user to modify or adjust different parameters, such as adjusting or otherwise changing different parameters of one or more component vectors described herein. In various embodiments, the interface may be used for training and reasoning purposes. For example, a user may evaluate a reconstruction on the interface and the values ​​associated with one or more features of a component vector to mark data that may be used for training purposes. In other embodiments, a user may also use such an interface to modify the displayed reconstruction.

[0054] The interface can include enabling the user to evaluate reconstruction and animation at a given time within the duration of the audio clip. As noted herein, because various embodiments can generate animations on small time clips (e.g., about 0.5 seconds) and can also not use previous information when generating animations, users can "skip" or otherwise move over different parts of the audio clip without loss in reconstruction quality due to lack of information from one or more previous frames. For example, a user can start watching an animation at a time point corresponding to 20 seconds in the clip, and then jump to a time point corresponding to 2 minutes in the clip, without loss of reconstruction quality due to not evaluating frames between 20 seconds and 2 minutes of the clip. Because different time points within an audio clip can be associated due to context vectors and will not have different outputs, various embodiments provide improved review and viewing by enabling skipping or otherwise tracking in an audio clip.

[0055] As described above, such an interface can be used as a type of post-processing at the time of reasoning, and in at least some embodiments, the post-processing can also be used to continue learning. For example, a user can view the generated animation playback through the interface, in which the animation of the character is presented. If the user thinks that the animation contains too much tension for the situation, the user can adjust one or more style selectors to reduce the tension or modify one or more other parameters, and cause one or more frames of the animation to be re-rendered. If the user detects that there is a bit of sadness in the character's speech that is not captured in the animation, the user can also adjust the setting. In some embodiments, the user may also be able to provide adjustments to specific feature points or facial components in the display as a type of style input. For example, a user can use a pointer to grab and move the position of the character's lips, and this information can be used as a style input for the re-rendering of the animation. Other changes, such as head movement, head tilt, eye movement or focus, or other such changes that can be conveyed by emotional or style input, can also be provided for the re-rendering (or updated rendering or synthesis) of the animation. Various other animation control parameters can also be specified through such an interface, which can affect the final rendering.

[0056] The various systems may also support retargeting. In retargeting, the motion of one character may be mapped to the motion of another character so that similar animations may be generated for similar emotions and / or styles. For example, retargeting may be applied to one or more (e.g., all) facial components, such as skin, jaw, tongue, eyeballs, etc. These facial components may be retargeted to more closely match or resemble the same facial component of a target or custom character. The interface may be further beneficial in the context of remapping where different characters may express emotions or styles in slightly different ways. A user may be able to load different characters into this interface and see how the retargeted rendering looks for that character, and then may modify one or more aspects or style of the motion or behavior of that particular character, or the type of character.

[0057] Figure 4An example training process 400 is shown that can be used to train a neural network for a task (such as facial animation) in accordance with at least one embodiment. It should be understood that for this process and other processes presented herein, there may be additional, fewer, or alternative operations performed in a similar or alternative order or at least partially in parallel within the scope of various embodiments, unless otherwise specifically stated. Further, while this example relates to facial animation, it should be understood that various other such tasks may benefit from aspects of such training processes that are well within the scope of various embodiments. In this process, audio and 4D image data (and / or other image data, such as 2D image data, 3D image data, etc.) are captured 402 or otherwise obtained for one or more actors speaking (including the same or different words or content) using different emotions and / or styles. In at least some embodiments, an actor speaking an utterance will be instructed to speak the utterance using at least a specified emotion or style, or some other parameter that can be modeled and represented using one or more component vectors, and further to "act" out that parameter during utterance in an attempt to capture image or video data representing realistic physical movement or behavior of one or more facial or body components during utterance with the specified parameters (e.g., emotion, style, etc.). This capture can be performed for any different number of emotions or phrases, and can include enough utterance instances for each emotion to enable accurate training of, for example, a deep neural network.

[0058] Reconstruction 404 of facial animation may be performed to provide a kind of ground truth data, wherein a character mesh or other representation is deformed based on the captured image data to provide a reference as to how facial components actually move or deform during the speaking of a particular utterance with a particular emotion and / or style. In addition, the ground truth information may correspond to the movement of individual facial features (e.g., lips, jaw, etc.) to determine whether the exact movement corresponds to the utterance associated with the audio data.

[0059] In at least one embodiment, a pre-trained converter model may be provided 406. The pre-trained converter model may be pre-trained on audio data that may include a variety of different languages ​​and may be used as an audio encoder within one or more pipelines. The pre-trained converter model may generate a set of encodings that are used as input to different layers of the converter model and / or decoders associated with the pre-trained converter model. In at least one embodiment, providing the pre-trained converter model includes locking or otherwise maintaining initial weights or parameters 408 of the pre-trained converter model so that these weights or parameters are not changed or updated during the training process.

[0060] At least a portion of the training data may then be provided 410 for use in training the deep neural network. This may include, for example, for each frame of the animation to be generated, a window of audio and component data (such as emotion data) for that frame. As mentioned, the audio and emotion data (and possibly any state data) may be used to generate a feature vector that may be provided as input to the network during training. As mentioned, a sequence-to-sequence mapping may be used to obtain a sufficiently long temporal context for the input data, which may be beneficial in generating physically or behaviorally accurate animations.

[0061] The neural network, after receiving and processing this data, can generate 412 an output vector corresponding to the linear combination of the output vectors of each frame of the pre-trained transformer model. For example, the input training data can be processed by one or more networks to extract one or more features, where these features are provided as input to the pre-trained transformer model. The pre-trained transformer model can then output vectors for each layer of the model, and these vectors are passed to the combination layer to linearly combine the vectors from each layer into an output vector.

[0062] Once generated, the output vector can be provided to a decoder 414, which can be part of a pre-trained model or can be a separate decoder, to update the weights of the decoder model based on the output vector. The decoder can then generate an output for use by one or more networks to generate 416 motion vectors for various points associated with one or more features of the output vector. For example, a neural network can generate a set of motion vectors (or vertices or deformation values, etc.) for one or more of one or more facial feature points of a character. This can include generating motion vectors for each facial feature point, or only generating motion vectors for those feature points that are subject to at least some amount of motion, as well as other such options. The network can then be deployed or otherwise provided 418 for inference, such as to generate facial animation data from input audio (and / or from one or more emotion vectors and / or one or more style vectors).

[0063] FIG5 illustrates an example process 500 for generating facial animation data using such a trained deep neural network, according to at least one embodiment. In this example, audio data is received 502, the audio data including at least some speech data for which a character is to be animated. In addition to the audio data, there may also be one or more component vectors including adjustments or settings to be applied to the speech animation (such as to the mood or style of the animation), which may be received 504 as additional input. The audio data and any received adjustment data may be provided 506 to the trained neural network, such as using a method similar to that described with respect to Figure 4The process described herein trains a neural network. The process may then receive 508 inferred motion vectors (or vertices, etc.) for each animation frame from the neural network, wherein those motion vectors indicate motion of feature points corresponding to a plurality of facial (or other body) components, such as a head, jaw, skin, eyeballs, or tongue, among other such options. These motion vectors are inferred to provide realistic animation for a character speaking the speech represented in the corresponding portion of the audio data, and to allow for separate modeling and behavior of these independent components within kinematic, structural, and / or other such constraints. In this example, these motion vectors may then be provided 510 to a renderer (or other such system, service, device, or process) for rendering one or more frames of facial animation by, for example, deforming one or more meshes of these facial components according to these motion vectors. The animation may then be provided 512 for purposes such as presentation or storage, among other such options.

[0064] Figure 5B An example process 520 for generating facial animation data using such a trained deep neural network is shown in accordance with at least one embodiment. In this example, audio data including speech of a speaker 522 is received. The audio data may include spoken phrases in various languages ​​and / or non-verbal spoken utterances, such as yelling, crying, sobbing, heavy breathing, etc. In at least one embodiment, the spoken phrases may be accompanied by a particular style, which may include singing, etc. Thus, the spoken phrases may refer to verbal expressions, and not necessarily only to speaking.

[0065] One or more neural networks may be incorporated with the transformer-based audio encoder to compute a weighted vector indicating a plurality of features associated with the audio data 524. The plurality of features may be determined by one or more neural networks, such as by a CNN that extracts feature information from the audio input before providing the feature information to the transformer-based audio encoder. For example, the CNN may learn which features of the input audio data are relevant to a given task, identify those features, extract those features, and then provide those features to the transformer-based audio encoder.

[0066] In at least one embodiment, one or more animation vectors may be calculated using the weighted vectors and one or more component vectors 526. The component vectors may correspond to, among other things, emotion or style vectors that modify different features of the output animation. For example, an emotion corresponding to "anger" may result in an output animation that causes the character to tighten their jaw and / or furrow their brows. Similarly, a style corresponding to singing may cause the animation to change its mouth position to project different words or keep its mouth position open longer to emphasize certain notes, etc. Based on the one or more animation vectors, a digital character representation may be rendered 528.

[0067] The systems and methods of the present disclosure can allow deep neural networks to effectively process human speech and generalize to different speakers. Such a process can also allow the network to discover changes in the training data that cannot be explained by audio alone, such as animations that may involve obvious emotional states, styles, or some other component driven. The three-way loss function presented in this article can also help ensure that the network remains temporally stable and responsive under animation, even with highly ambiguous training data. Such a process can generate expressive 3D facial movements from audio in real time or near real time (e.g., without significant delay) and with low latency. In order to remain independent of the details of downstream animation systems, such a system can output the per-frame positions of the control vertices of a fixed topology facial mesh. If compression, rendering, or editability require, alternative encodings such as blend shapes or nonlinear rigs can be introduced in later pipeline stages. An example network can be trained using three to five minutes of high-quality continuous footage obtained using traditional vision-based performance capture methods. It has been observed that such a process successfully models not only the speaking style of a single actor, but also the speaking styles of other speakers from different genders, accents, or languages. This flexibility can be used for various applications or operations, such as those involving in-game dialogue, low-cost localization, virtual reality, augmented reality, mixed reality, augmented reality and telepresence, and other such options. Even in film, such an approach can prove useful in accommodating small script changes.

[0068] An example and non-limiting deep neural network (including at least a CNN) consists of one dedicated layer, ten convolutional layers, two fully connected layers, and at least a transformer model having about six layers, such as Figure 2 and Figure 3 As shown in , the converter model can be divided into three conceptual parts. In at least one embodiment, the system and method include a pipeline that receives input audio, processes the input audio through one or more CNN layers, processes the output of the CNN layer through one or more converter layers, processes the output of the converter layer through a linear combination layer, and then provides the output to the pronunciation layer. The various outputs from different layers can be routed to various other layers, as noted herein, such as providing each output from each converter layer to a combination layer, and therefore, various embodiments can have various output vectors from various different layers, which are processed, combined, modified, etc. The data of the input audio window can be fed to a resonance peak analysis network to generate a time-varying sequence of speech features that will subsequently drive pronunciation. Through training, the convolutional layer learns to extract short-term features related to facial animation, such as pitch, strong tone, and specific phonemes. Their abstract time-varying representation can be the output of the final convolutional layer.

[0069] The result can be fed to a converter including an encoder. The encoder can be a pre-trained audio encoder trained on a variety of different audio data (such as multiple languages, mixed languages, etc.). The parameters of the audio encoder can be frozen, while other parameters of the system are changed during training. In this example, the pronunciation network can also form part of a deep neural network consisting of five further convolutional layers, which analyze the time evolution of features and ultimately determine a single abstract feature vector describing the facial gesture at the center of the audio window. As an auxiliary input, the pronunciation network accepts (learned) descriptions of emotional states and / or styles (which can be called component vectors) to disambiguate between different facial expressions and speaking styles. Emotional states can be represented as E-dimensional vectors, either alone or together with style data, which are directly concatenated to the output of each layer in the pronunciation network, so that subsequent layers can change their behavior accordingly.

[0070] In at least one embodiment, the network architecture may include a CNN (e.g., CNN 302) in the analysis network 208, which has a total of seven 1D (along the time axis) convolutions. The convolution kernel sizes of the seven layers may be 10, 3, 3, 3, 3, 2, 2, respectively. The strides of the convolutions of these seven layers may be 5, 2, 2, 2, 2, 2, 2. The output of the convolution may have a shape of 25 (time dimension) x512 (feature dimension), which is passed to the feature projection layer to change its shape to 25x768. This 25x768-shaped output may be fed to a converter layer (e.g., layer 308) (e.g., which may include six or twelve layers, or any other reasonable number of layers), which also outputs features with a 25x768 shape from each layer. These outputs may then be passed to a linear combination layer (e.g., layer 312).

[0071] The example pronunciation network can output a set of 256+E+S abstract features that together represent the desired facial pose - for example, E is used for the dimension of the emotion vector and S is used for the dimension of the style vector. In at least one embodiment, these features can be fed to the output network to produce the final 3D position of the set of control vertices in the tracking mesh. In addition, in various embodiments, the output network calculates PCA coefficients for the face mesh (e.g., 140 in total), the tongue mesh (e.g., 10 in total), and the rigid transformation coefficients for the eyeballs (e.g., 4 in total) and / or the jaw (e.g., 15 in total). The output network can be implemented as a pair of fully connected layers that perform a simple linear transformation on the data. The first layer maps the input feature set to the weights of the linear basis, and the set of second layers calculates the final PCA coefficients for the face and tongue, the rotation values ​​for the eyeballs, and the translation displacements for the jaw and head. If the output network is calculating the 3D position of the vertex as output, the second layer can be initialized to, for example, 150 pre-computed PCA components, which together explain about 99.9% of the variance seen in the training data. In one or more embodiments, the two linear transformation layers of the output network may have an architecture such that the first transformation layer transforms features of size 256+E+S into features of shape 169. The second transformation layer then takes an input of shape 169 and outputs coefficients of shape 169 (140+10+4+15). The weights of the transformation layers may be randomly initialized.

[0072] The main input to such a network is a speech audio signal, which can be converted to a format such as 16kHz mono audio before feeding the audio to the network. The volume of each vocal track can be normalized to use the full [-1, +1] dynamic range, but such a system may or may not employ other kinds of processing, such as dynamic range compression, noise reduction, or pre-emphasis tone filters.

[0073] In one example implementation, 520ms worth of audio is used as input, e.g., as 260ms of past and future samples about the desired output pose. This value is chosen to capture relevant effects, such as phoneme co-articulation, without providing too much data that could lead to overfitting.

[0074] Inferring facial animation from speech can be an inherently ambiguous problem, since the same sound can be produced with very different facial expressions. This is especially true for eyes and eyebrows, since they have no direct causal relationship to sound production. Such ambiguity is also problematic for deep neural networks, since the training data will inevitably contain cases where nearly identical audio inputs are expected to produce very different output gestures. If the network has nothing else to process besides the audio, it will learn to output the statistical mean of conflicting outputs.

[0075] An example method to resolve this ambiguity is to introduce at least an auxiliary input into the network. A small amount of additional potential data can be associated with each training sample so that the network has enough information to unambiguously infer the correct output pose. This additional data can encode all relevant aspects of the animation in the neighborhood of a given training sample that cannot be inferred from the audio itself, including different facial expressions and co-articulation patterns. The auxiliary input can include predefined labels and can at least represent the emotional state of the actor. In addition to resolving ambiguity in the training data, such an auxiliary input can also be very useful for reasoning because it enables the system to mix and match different emotional states with a given soundtrack to provide powerful control over the resulting animation.

[0076] In addition to or as an alternative to relying on predefined labels, the system according to at least one embodiment can also adopt a data-driven approach, in which the network automatically learns a concise representation of the style as part of the training process. This allows the system to extract meaningful emotional states even from continuous shots within the character, as long as there is a sufficient range of emotions. In at least one embodiment, the style state can be represented by an S-dimensional vector, where S is an adjustable parameter that can be set to a value such as but not limited to 16 or 24, and the components are initialized to random values ​​drawn from a Gaussian distribution. Such a vector can be assigned to each training sample, and the matrix storing these latent variables is referred to as a style database in this article. The style data can be attached to the activation lists of all layers of the pronunciation network, which can make it part of the computational graph of the loss function, and as a trainable parameter, it can be updated during back propagation with the network weights. In this example, the dimension S is a compromise between the two effects. If S is too low, the style cannot disambiguate the changes in the training data, resulting in a weak audio response. If S is too high, the style may become too specialized to be used for general reasoning.

[0077] By design, the information provided by audio can be limited to short-term effects within, for example, 520ms intervals. Therefore, a natural way to prevent styles from containing repetitive information is to prohibit them from containing short-term changes. Focusing styles on long-term effects may also be desirable for inference, as the network may be expected to produce plausible animations even when the emotional state is held fixed. This requirement can be expressed by introducing a dedicated regularization term in the loss function to penalize fast changes in the style database, which results in incremental smoothing of the emotional state during training. A potential limitation to such an approach is that aspects such as blinks and eye movements may not be modeled correctly, as they are not correlated with the audio and cannot be represented using slowly changing emotional states.

[0078] In an embodiment, emotion and style states may be attached to all layers of the pronunciation network to help significantly improve results in practice, since emotion and style states can control animations at multiple levels of abstraction, and higher levels of abstraction may be more difficult to learn. Connecting to earlier layers may provide subtle control over subtle animation features (such as co-articulation), while connecting to later layers may provide more direct control over output poses. Early stages of training may focus on the latter, while later stages may focus on the former once individual poses are reasonably well represented.

[0079] In one method of training a deep neural network, an unstructured mesh with texture and optical flow data can be reconstructed from, for example, nine images captured for each frame. A template mesh of fixed topology created using a separate photogrammetry pipeline prior to capture can be projected onto the unstructured mesh and associated with the optical flow. The performance of the template mesh can be tracked, and any problems are semi-automatically fixed, such as by a tracking artist in software. Several key vertices of the tracking mesh can be used to stabilize the position and orientation of the head. Finally, the vertex positions of the mesh can be derived for each frame in the shot. These positions (or, more precisely, deltas from a neutral pose) can be the target output of the network when given an audio window during training.

[0080] For each actor, the training set can be composed of at least two parts: a pangram and in-role material. In general, the quality of reasoning can increase with the growth of the training set, but due to the cost of capturing high-quality training data, a small training set may be highly desirable. In at least one embodiment, it is empirically determined that about 3 to 5 minutes per actor represents the actual sweet spot. The pangram collection can try to cover a set of possible facial movements during normal speech in a given target language (such as English). The actor speaks one to three pangrams (e.g., sentences designed to contain as many different phonemes as possible) in several different emotional tones to provide good coverage of the range of expression. The in-role material collection can take advantage of the fact that the performance of the actor's role is often seriously deviated in terms of emotion and range of expression due to different dramatic and narrative reasons. In the case of a movie or game product, the material can be composed of a preliminary version of the script. Only shots that are considered to support different aspects of the role are selected to ensure that even if the reasoning is not perfect, or even if completely novel or voice actions that do not conform to the role are encountered, the trained network can also produce outputs that remain in the role.

[0081] Given the potentially ambiguous nature of the training data, one can strive to define a meaningful loss function to be optimized. In at least one embodiment, a specialized loss function consisting of three different terms can be used: a position term to ensure that the overall position of each output vertex is roughly correct; a motion term to ensure that the vertices exhibit the right kind of movement under animation; and a regularization term to discourage the style database from containing short-term changes.

[0082] In practice, simultaneous optimization of multiple loss terms can be difficult because these terms can have wildly different magnitudes during training and their balance can change in unpredictable ways. One solution is to associate predefined weights with each term to ensure that the optimization does not overlook any of them. However, choosing optimal values ​​for the weights can be a tedious process of trial and error, which may need to be repeated whenever the training set changes. To overcome these problems, a normalization scheme can be used that automatically balances the loss terms with respect to their relative importance. Thus, an equal amount of work can be invested in optimizing each term so that no additional weights need to be specified.

[0083] One error metric that can be used is the difference between the desired output y and the output produced by the network For a given training sample x, this can be expressed using the position term P(x):

[0084]

[0085] Here, N represents the total number of output features including the 3D positions of the skin / tongue mesh vertices, the rotation values ​​of the eyeballs, and the translation displacement of the jaw / head, and y (i) It means y=(y (1) ,y (2) ,...,y (N) ). By way of example, N would be 61019 (e.g., 60000 for the face mesh, 1000 for the tongue mesh, 4 for the rotation of the eyes, and 15 for the translation of the jaw).

[0086] Even if the position term ensures that the output of the network is approximately correct at any given moment, it may not be enough to produce high-quality animations in all cases. Observe that training the network with only the position term leads to a large amount of temporal instability, and the response to individual phonemes is generally weak. Therefore, the network can also be optimized in terms of vertex motion: a given output vertex should move only if it also moves in the training data, and it only moves at the right time. Therefore, the system can address vertex motion as part of the loss function.

[0087] One approach for training a neural network is to iterate over minibatches of training data, where each minibatch consists of B randomly selected training examples x1,x2,...,x B To account for vertex motion, we plot the samples as B / 2 time pairs, each consisting of two adjacent frames. The operator m[·] can be defined as the finite difference between pairs of frames, which allows the motion term M(x) to be defined as:

[0088]

[0089] In this equation, the factor of 2 occurs because M is evaluated once per time pair.

[0090] Additionally, it can be beneficial to ensure that the network correctly attributes short-term effects to the audio signal and long-term effects to emotional states. One approach could use the same finite difference operator as above to define the regularization term for the emotion / style database:

[0091]

[0092] Here, e( i )(x) represents the i-th component of the training sample x stored by the emotion database. It can be noted that this definition does not explicitly prohibit the emotion / style database from containing short-term variations - instead it discourages too much variation in the average. This may be significant in at least some instances, because the training data may occasionally contain legitimate short-term changes in emotional state, and the network may not want to incorrectly try to interpret them based on the audio signal.

[0093] The caveat of Equation 3 is that R'(x) can be made arbitrarily close to zero by simply decreasing the magnitude of e(x) while increasing the corresponding weights in the network. The idea behind batch normalization is to normalize R relative to the observed magnitude of e(x) by ￠ (x) to remove the worthless solution:

[0094]

[0095] To balance these three loss terms, one approach is to exploit the properties of the Adam (or other) optimization method used to train the network. In practice, Adam updates the weights of the network according to the gradient of the loss function, which is normalized in a component-by-component manner according to a long-term estimate of its second raw moment. This normalization makes the training resistant to differences in the magnitude of the loss function, but this is only true for the loss function as a whole - not for individual terms. One approach is to perform a similar normalization for each loss term separately. Using the position term as an example, the second raw moment of P(x) can be estimated for each mini-batch and a moving average v is maintained across consecutive mini-batches. P t , as can be given by the following formula:

[0096]

[0097] Here, t represents the mini-batch index and β is a decay parameter of the moving average that can be set to a value such as, but not limited to, 0.99. The system can be initialized and correct this estimate to take the startup bias into account to obtain Then we can calculate the mean value P(x) on the current mini-batch and use Normalize the value:

[0098]

[0099] In Equation 6, ∈ can be set to, for example, 10- 8 The value of is such that a small constant is avoided to avoid division by zero. Repeating equations 7 and 8 for M(·) and R(·), the final loss function can be expressed as the sum of the terms l=l P +l M +l R In some embodiments, the importance of the loss term can be further fine-tuned by attaching weights.

[0100] Similarly, the loss during training can also be associated with lip distance regularization and body-based stability regularization, as described herein. For example, key points can be selected on the lips (e.g., upper and lower lips), and then the minimum difference between the predicted distance and the true distance between the key points. The loss can have a higher weight when the true lip distance is small (e.g., when the mouth is closed). In addition, if the sound volume is low, a loss can be added to penalize large motions between adjacent animation frames. Therefore, the final loss can be further fine-tuned to include a loss of one or both of lip distance or body-based stability.

[0101] In at least one embodiment, random time shifts can be used for training samples to improve temporal stability and reduce overfitting. Each time a small batch is presented to the network, the input audio window can be randomly shifted by up to 16.6ms (±0.5 frames at 30FPS) in either direction. To compensate, the same shift can be applied to the desired output pose by linear interpolation. The two training samples in a time pair can be shifted by the same amount, where different random shift amounts are used for different pairs. In some embodiments, cubic interpolation can be used on the output instead of or in addition to linear interpolation.

[0102] To improve generalization and avoid overfitting, multiplicative noise can be applied to the inputs of individual convolutional layers. The noise can have the same amplitude for each layer and can be applied on a per-feature map basis so that all activations for a given feature map are multiplied by the same factor. The same noise can be applied to pairs of training examples to obtain meaningful motion terms. One formula for this noise is 1.4 N(0,1) Besides time shifting of the input / output and multiplicative noise inside the network, there may be no other types of noise or augmentation applied to the training samples. However, some methods may perform multiple operations such as adjusting volume, adding reverberation (both long and short), performing time stretching and pitch shifting, applying nonlinear distortion, random equalization, and scrambling phase information, etc.

[0103] Once trained, the deep neural network can be evaluated at any point in time by selecting an appropriate audio window, resulting in facial animation at a desired frame rate. The latency of such an approach may depend at least in part on the audio window size, which may extend into time periods in the past and / or future. Co-articulation may set a lower limit for look-ahead; it has been observed that during training, look-ahead may be limited to values ​​such as 100ms with little degradation in quality, even though some co-articulation effects may be longer. Shortening the look-ahead much further than this may in some cases result in a rapid drop in perceived responsiveness, so a practical lower limit for the latency of one embodiment may be set to about 100ms.

[0104] When inferring facial poses for novel audio, the network can be provided with an emotional state vector and / or a style vector as an auxiliary input, which can also be part of a single emotion vector. As part of training, the network can learn a vector (e.g., a latent E-dimensional vector) for each training sample, and this emotion database can be used to obtain a robust emotion vector that can be used during inference.

[0105] During training, the network can attempt to separate latent information (e.g., everything that cannot be inferred from the audio alone) into a mood / style database. However, this decomposition results in some amount of crosstalk between pronunciation and total expression. In practice, many of the learned mood / style vectors may only apply to the neighborhood of their corresponding training frames and are not necessarily useful for general reasoning. In at least one embodiment, the process can use a three-operation process to mine robust mood / style vectors. The problem experienced in many learned mood vectors is that they no longer emphasize the movement of the mouth: when such vectors are used as constant input when performing reasoning on new audio, the apparent movement of the mouth can be suppressed. One approach is to select several audio windows from a validation set containing bilabial sounds and several audio windows from a validation set containing vowels, for which the mouth should be closed and opened, respectively. The mood / style database can then be scanned to find vectors that exhibit the desired behavior for all selected windows. This preliminary elimination is performed for character 1 generated in 100 candidate mood vectors for further consideration, and this response may vary with different mood vectors.

[0106] The second operation in this example culling process is to play the validation track and check the facial motion inferred by each of the candidate emotion / style vectors. At this stage, vectors that result in soft or false, unnatural motion can be discarded, indicating that the vector may be contaminated by short-term effects. This stage narrows the set to 86 candidate emotion vectors for character 1. As the third and final operation in this example, reasoning can be run on several seconds of audio from different speakers and vectors of silence or unnatural responses are eliminated. For character 1, this operation leaves 33 emotion vectors.

[0107] The output of the network can be examined for several novel audio clips, which have each remaining emotion / style vector and a "semantic meaning" assigned to each of them (e.g., "neutral," "entertained," "surprised," etc.), which depends at least in part on factors such as the emotional state they convey. Which semantic emotions are retained can depend on the training material, and if the training data does not contain enough such material that can be generalized to novel audio, it may be impossible to extract, for example, a "happy" emotion. Even after removing all but the best-performing emotion vectors, there is still a large amount of variation to choose from. It has been observed that emotion vectors mined in this way perform well under interpolation, for example, extending from one emotion vector to another tends to produce results that look natural. Therefore, emotional states can be changed during reasoning based on high-level information from the game engine or by manually setting keyframes.

[0108] The resulting facial animation can be highly stable. The main sources of this temporal stability can include motion terms l M and time-shift enhancement, but even with these techniques a small amount of jitter may remain, e.g. on the 4ms timescale in the lip region for some inputs. This can be caused by aliasing between adjacent audio frames around features such as stops and pops. This can be mitigated at least in part using at least some amount of integration: evaluate the network twice for a given animation frame, separated in time (e.g., 4ms), and average the predictions.

[0109] As mentioned above, this approach can also support retargeting. When training the model, the output network can become specialized for a specific mesh. For many operations, it may be desirable to use audio input to drive several different meshes. The methods discussed herein can support retargeting of morphs, or transfer of morphing behavior between characters or for the same character at different stages of life, etc.

[0110] As discussed, aspects of the various methods presented herein can be sufficiently lightweight to be performed in real time on a device (such as a client device, such as a personal computer or a game console). Such processing can be performed on or for content (such as streaming data or other content received over at least one network) generated on or received by the client device or received from an external source. In some instances, the processing and / or determination of this content can be performed by one of these other devices, systems, or entities and then provided to the client device (or another such recipient) for presentation or another such use.

[0111] As an example, Figure 6An example network configuration 600 is shown that can be used to provide, generate, modify, encode, process and / or transmit image data or other such content. In at least one embodiment, a client device 602 can generate or receive data for a session using components of a control application 604 on the client device 602 and data stored locally on the client device. In at least one embodiment, a content application 624 executed on a server 620 (e.g., a cloud server or edge server) can initiate a session associated with at least one client device 602, such as can utilize a session manager and user data stored in a user database 636, and can cause content (such as one or more digital assets (e.g., object representations)) from an asset repository 634 to be determined by a content manager 626. The content manager 626 can work with an audio-to-face module 628 to generate or synthesize new objects, digital assets, or other such content to be provided for presentation via the client device 602. In at least one embodiment, this audio-to-face module 628 can use one or more neural networks or machine learning models that can be trained or updated using a training module 632 or system on or in communication with the server 620. This may include training and / or using a diffusion model 630 to generate content tiles that can be used by the audio to face module 628, for example to apply a non-repeating texture to an area of ​​the environment where image or video data will be presented via the client device 602, which may include using one or more renderers 630. At least a portion of the generated content may be transmitted to the client device 602 using an appropriate transmission manager 622 for transmission via download, streaming, or another such transmission channel. An encoder may be used to encode and / or compress at least some of the data before sending it to the client device 602. In at least one embodiment, a client device 602 receiving such content may provide the content to a corresponding control application 604, which may also or alternatively include a graphical user interface 610, an audio to face manager 612, and a renderer 614 for providing, synthesizing, modifying, or using content for presentation on or by the client device 602 (or other purposes). The decoder may also be used to decode data received over one or more networks 640 for presentation via the client device 602, such as presentation of images or video content through a display 606 and presentation of audio (such as sounds and music) through at least one audio playback device 608 (such as a speaker or headphones). In at least one embodiment, at least some of the content may already be stored on, rendered on, or accessible to the client device 602, such that transmission over the network 640 is not required for at least that portion of the content, such as the content may have been previously downloaded or stored locally on a hard drive or optical disk.In at least one embodiment, a transport mechanism such as data streaming may be used to transmit the content from the server 620 or user database 636 to the client device 602. In at least one embodiment, at least a portion of the content may be obtained, enhanced, and / or streamed from another source such as a third-party service 660 or other client device 650, which may also include a content application 662 for generating, enhancing, or providing the content. In at least one embodiment, multiple computing devices or multiple processors within one or more computing devices (such as a combination of a CPU and a GPU) may be used to perform portions of this functionality.

[0112] In this example, these client devices may include any appropriate computing devices, such as desktop computers, laptops, set-top boxes, streaming devices, game consoles, smartphones, tablet computers, VR headsets, AR goggles, wearable computers, or smart TVs. Each client device may submit a request via at least one wired or wireless network, which may include the Internet, Ethernet, a local area network (LAN), or a cellular network, as well as other such options. In this example, these requests may be submitted to an address associated with a cloud provider, which may operate or control one or more electronic resources in a cloud provider environment, such as a data center or a server farm. In at least one embodiment, the request may be received or processed by at least one edge server located at the edge of the network and outside of at least one security layer associated with the cloud provider environment. In this way, latency can be reduced by enabling client devices to interact with servers that are closer, while also improving the security of resources in the cloud provider environment.

[0113] In at least one embodiment, such a system can be used to perform graphics rendering operations. In other embodiments, such a system can be used for other purposes, such as for providing image or video content to test or verify autonomous machine applications, or for performing deep learning operations. In at least one embodiment, such a system can be implemented using an edge device, or can incorporate one or more virtual machines (VMs). In at least one embodiment, such a system can be implemented at least in part in a data center or at least in part using cloud computing resources.

[0114] Reasoning and training logic

[0115] Fig. 7A Inference and / or training logic 715 is shown for performing reasoning and / or training operations associated with one or more embodiments. Fig. 7A and / or Figure 7B Details regarding the inference and / or training logic 715 are provided.

[0116] In at least one embodiment, the inference and / or training logic 715 may include, but is not limited to, code and / or data storage 701 for storing forward and / or output weights and / or input / output data, and / or other parameters for configuring neurons or layers of a neural network trained and / or used for inference in aspects of one or more embodiments. In at least one embodiment, the training logic 715 may include or be coupled to code and / or data storage 701 for storing graph code or other software to control timing and / or sequence, where weights and / or other parameter information is loaded to configure logic, including integer and / or floating point units (collectively referred to as arithmetic logic units (ALUs)). In at least one embodiment, the code (such as graph code) loads weights or other parameter information into a processor ALU based on the architecture of the neural network to which the code corresponds. In at least one embodiment, the code and / or data storage 701 stores input / output data during training and / or inference using aspects of one or more embodiments and / or weight parameters during forward propagation of each layer of a neural network trained or used in conjunction with one or more embodiments. In at least one embodiment, any portion of code and / or data storage 701 may be included within other on-chip or off-chip data storage, including a processor's L1, L2, or L3 cache or system memory.

[0117] In at least one embodiment, any portion of code and / or data storage 701 may be internal or external to one or more processors or other hardware logic devices or circuits. In at least one embodiment, code and / or data storage 701 may be cache memory, dynamic random access memory ("DRAM"), static random access memory ("SRAM"), non-volatile memory (e.g., flash memory), or other storage. In at least one embodiment, the choice of whether code and / or data storage 701 is internal or external to a processor, for example, or consists of DRAM, SRAM, flash memory, or some other storage type, may depend on the available storage space on-chip or off-chip, the latency requirements of the training and / or inference functions being performed, the batch size of the data used in inference and / or training of the neural network, or some combination of these factors.

[0118] In at least one embodiment, the inference and / or training logic 715 may include, but is not limited to, code and / or data storage 705 for storing reverse and / or output weights and / or input / output data corresponding to neurons or layers of a neural network trained and / or used for inference in aspects of one or more embodiments. In at least one embodiment, during training and / or inference using aspects of one or more embodiments, the code and / or data storage 705 stores weight parameters and / or input / output data for each layer of a neural network trained or used in conjunction with one or more embodiments during back propagation of input / output data and / or weight parameters. In at least one embodiment, the training logic 715 may include or be coupled to code and / or data storage 705 for storing graph code or other software to control timing and / or sequence, wherein weights and / or other parameter information is loaded to configure logic including integer and / or floating point units (collectively referred to as arithmetic logic units (ALUs)). In at least one embodiment, code (such as graph code) loads weights or other parameter information into a processor ALU based on the architecture of the neural network to which the code corresponds. In at least one embodiment, any portion of code and / or data storage 705 may be included with other on-chip or off-chip data storage, including the processor's L1, L2, or L3 cache or system memory. In at least one embodiment, any portion of code and / or data storage 705 may be internal or external to one or more processors or other hardware logic devices or circuits. In at least one embodiment, code and / or data storage 705 may be cache memory, DRAM, SRAM, non-volatile memory (e.g., flash memory), or other storage. In at least one embodiment, the choice of whether code and / or data storage 705 is internal or external to the processor, for example, whether it is composed of DRAM, SRAM, flash memory, or some other storage type, depends on whether the available storage is on-chip or off-chip, the latency requirements of the training and / or inference functions being performed, the data batch size used in the inference and / or training of the neural network, or some combination of these factors.

[0119] In at least one embodiment, code and / or data storage 701 and code and / or data storage 705 may be separate storage structures. In at least one embodiment, code and / or data storage 701 and code and / or data storage 705 may be the same storage structure. In at least one embodiment, code and / or data storage 701 and code and / or data storage 705 may be partially the same storage structure and partially separate storage structures. In at least one embodiment, any portion of code and / or data storage 701 and code and / or data storage 705 may be included with other on-chip or off-chip data storage, including the processor's L1, L2, or L3 cache or system memory.

[0120] In at least one embodiment, the inference and / or training logic 715 may include, but is not limited to, one or more arithmetic logic units (“ALUs”) 710 (including integer and / or floating point units) for performing logical and / or mathematical operations based at least in part on or as directed by training and / or inference code (e.g., graph code), the results of which may produce activations (e.g., output values ​​from a layer or neuron within a neural network) stored in activation storage 720, which is a function of input / output and / or weight parameter data stored in code and / or data storage 701 and / or code and / or data storage 705. In at least one embodiment, activations are performed in response to executing instructions or other code, linear algebra and / or matrix-based mathematics performed by ALU 710 to generate activations stored in activation storage 720, where weight values ​​stored in code and / or data storage 705 and / or in code and / or data storage 701 are used as operands with other values, such as bias values, gradient information, momentum values, or other parameters or hyperparameters, any or all of which may be stored in code and / or data storage 705 or code and / or data storage 701 or other on-chip or off-chip storage.

[0121] In at least one embodiment, one or more ALUs 710 are included in one or more processors or other hardware logic devices or circuits, while in another embodiment, one or more ALUs 710 may be outside a processor or other hardware logic device or circuit that uses them (e.g., a coprocessor). In at least one embodiment, one or more ALUs 710 may be included within an execution unit of a processor, or otherwise included in a group of ALUs accessible by an execution unit of a processor, which may be within the same processor or distributed between different processors of different types (e.g., a central processing unit, a graphics processing unit, a fixed function unit, etc.). In at least one embodiment, code and / or data storage 701, code and / or data storage 705, and activation storage 720 may be on the same processor or other hardware logic device or circuit, while in another embodiment, they may be in different processors or other hardware logic devices or circuits or some combination of the same and different processors or other hardware logic devices or circuits. In at least one embodiment, any portion of activation storage 720 may be included with other on-chip or off-chip data storage, including the processor's L1, L2, or L3 cache or system memory. Additionally, inference and / or training code may be stored with other code accessible to a processor or other hardware logic or circuitry and may be retrieved and / or processed using the processor's fetch, decode, schedule, execute, exit, and / or other logic circuitry.

[0122] In at least one embodiment, activation storage 720 may be cache memory, DRAM, SRAM, non-volatile memory (e.g., flash memory), or other storage. In at least one embodiment, activation storage 720 may be completely or partially internal or external to one or more processors or other logic circuits. In at least one embodiment, activation storage 720 may be selected to be internal or external to a processor, for example, or include DRAM, SRAM, flash memory, or other storage types, depending on the storage available on-chip or off-chip, the latency requirements for performing training and / or inference functions, the batch size of data used in inferencing and / or training neural networks, or some combination of these factors. In at least one embodiment, Fig. 7A The inference and / or training logic 715 shown in FIG. 7 may be used in conjunction with an application specific integrated circuit (“ASIC”), such as the ASIC from Google. Processing unit from Graphcore TM Inference Processing Unit (IPU) from Intel Corp (e.g., "LakeCrest") processor. In at least one embodiment, Fig. 7A The illustrated inference and / or training logic 715 may be used in conjunction with central processing unit (“CPU”) hardware, graphics processing unit (“GPU”) hardware, or other hardware such as a field programmable gate array (“FPGA”).

[0123] Figure 7B Inference and / or training logic 715 is shown in accordance with at least one or more embodiments. In at least one embodiment, the reasoning and / or training logic 715 may include, but is not limited to, hardware logic in which computing resources are dedicated or otherwise uniquely used in conjunction with weight values ​​or other information corresponding to one or more layers of neurons within a neural network. In at least one embodiment, Figure 7B The inference and / or training logic 715 shown in FIG. 7 may be used in conjunction with an application specific integrated circuit (ASIC), such as the ASIC from Google. Processing unit from Graphcore TM Inference Processing Unit (IPU) from Intel Corp (e.g., "LakeCrest") processor. In at least one embodiment, Figure 7BThe reasoning and / or training logic 715 shown in can be used in conjunction with central processing unit (CPU) hardware, graphics processing unit (GPU) hardware, or other hardware (e.g., field programmable gate array (FPGA)). In at least one embodiment, the reasoning and / or training logic 715 includes, but is not limited to, code and / or data storage 701 and code and / or data storage 705, which can be used to store code (e.g., graph code), weight values, and / or other information, including bias values, gradient information, momentum values, and / or other parameter or hyperparameter information. Figure 7B In at least one embodiment shown in , each of code and / or data storage 701 and code and / or data storage 705 is associated with a dedicated computing resource (e.g., computing hardware 702 and computing hardware 706), respectively. In at least one embodiment, each of computing hardware 702 and computing hardware 706 includes one or more ALUs that perform mathematical functions (e.g., linear algebraic functions) solely on the information stored in code and / or data storage 701 and code and / or data storage 705, respectively, and the results of performing the functions are stored in activation storage 720.

[0124] In at least one embodiment, each of the code and / or data stores 701 and 705 and the corresponding computing hardware 702 and 706 corresponds to a different layer of the neural network, such that activations from one "storage / compute pair 701 / 702" of the code and / or data store 701 and computing hardware 702 are provided as inputs to the next "storage / compute pair 705 / 706" of the code and / or data store 705 and computing hardware 706, so as to reflect the conceptual organization of the neural network. In at least one embodiment, each storage / compute pair 701 / 702 and 705 / 706 can correspond to more than one neural network layer. In at least one embodiment, additional storage / compute pairs (not shown) can be included in the inference and / or training logic 715 after or in parallel with the storage / compute pairs 701 / 702 and 705 / 706.

[0125] Data Center

[0126] Figure 8 An example data center 800 is shown in which at least one embodiment may be used. In at least one embodiment, the data center 800 includes a data center infrastructure layer 810, a framework layer 820, a software layer 830, and an application layer 840.

[0127] In at least one embodiment, Figure 8As shown, the data center infrastructure layer 810 may include a resource coordinator 812, grouped computing resources 814, and node computing resources ("node CRs") 816(1)-816(N), where "N" represents any positive integer. In at least one embodiment, the node CRs 816(1)-816(N) may include, but are not limited to, any number of central processing units ("CPUs") or other processors (including accelerators, field programmable gate arrays (FPGAs), graphics processors, etc.), memory devices (e.g., dynamic read-only memories), storage devices (e.g., solid-state drives or disk drives), network input / output ("NWI / O") devices, network switches, virtual machines ("VMs"), power modules and cooling modules, etc. In at least one embodiment, one or more of the node CRs 816(1)-816(N) may be a server having one or more of the above computing resources.

[0128] In at least one embodiment, the grouped computing resources 814 may include a separate grouping (not shown) of node CRs housed in one or more racks, or many racks (also not shown) housed in data centers at various geographic locations. The separate grouping of node CRs within the grouped computing resources 814 may include computing, networks, memory, or storage resources that can be configured or allocated to support groupings of one or more workloads. In at least one embodiment, several node CRs including a CPU or processor may be grouped in one or more racks to provide computing resources to support one or more workloads. In at least one embodiment, one or more racks may also include any number of power modules, cooling modules, and network switches in any combination.

[0129] In at least one embodiment, resource coordinator 812 may configure or otherwise control one or more node CRs 816(1)-816(N) and / or grouped computing resources 814. In at least one embodiment, resource coordinator 812 may include a software design infrastructure ("SDI") management entity for data center 800. In at least one embodiment, resource coordinator 812 may include hardware, software, or some combination thereof.

[0130] In at least one embodiment, Figure 8As shown, the framework layer 820 includes a job scheduler 822, a configuration manager 824, a resource manager 826, and a distributed file system 828. In at least one embodiment, the framework layer 820 may include a framework that supports software 832 of the software layer 830 and / or one or more applications 842 of the application layer 840. In at least one embodiment, the software 832 or the application 842 may include a web-based service software or application, such as a service or application provided by Amazon Web Services, Google Cloud, and Microsoft Azure. In at least one embodiment, the framework layer 820 may be, but is not limited to, a free and open source software network application framework, such as Apache SparkTM (hereinafter referred to as "Spark") that can utilize the distributed file system 828 for large-scale data processing (e.g., "big data"). In at least one embodiment, the job scheduler 832 may include a Spark driver to facilitate scheduling of workloads supported by various layers of the data center 800. In at least one embodiment, the configuration manager 824 may be able to configure different layers, such as the software layer 830 and the framework layer 820 including Spark and a distributed file system 828 for supporting large-scale data processing. In at least one embodiment, the resource manager 826 can manage cluster or group computing resources mapped to or allocated to support the distributed file system 828 and the job scheduler 822. In at least one embodiment, the cluster or group computing resources can include group computing resources 814 on the data center infrastructure layer 810. In at least one embodiment, the resource manager 826 can coordinate with the resource coordinator 812 to manage these mapped or allocated computing resources.

[0131] In at least one embodiment, the software 832 included in the software layer 830 may include software used by at least a portion of the node CRs 816(1)-816(N), the grouped computing resources 814, and / or the distributed file system 828 of the framework layer 820. The one or more types of software may include, but are not limited to, Internet web page search software, email virus scanning software, database software, and streaming video content software.

[0132] In at least one embodiment, one or more applications 842 included in the application layer 840 may include one or more types of applications used by at least a portion of the node CRs 816(1)-816(N), the grouped computing resources 814, and / or the distributed file system 828 of the framework layer 820. The one or more types of applications may include, but are not limited to, any number of genomics applications, cognitive computing, and machine learning applications, including training or inference software, machine learning framework software (e.g., PyTorch, TensorFlow, Caffe, etc.), or other machine learning applications used in conjunction with one or more embodiments.

[0133] In at least one embodiment, any of the configuration manager 824, resource manager 826, and resource coordinator 812 can implement any number and type of self-modification actions based on any number and type of data acquired in any technically feasible manner. In at least one embodiment, the self-modification actions can relieve a data center operator of the data center 800 from making potentially bad configuration decisions and can avoid underutilized and / or poorly performing portions of the data center.

[0134] In at least one embodiment, the data center 800 may include tools, services, software, or other resources to train one or more machine learning models or use one or more machine learning models to predict or infer information according to one or more embodiments described herein. For example, in at least one embodiment, a machine learning model may be trained by calculating weight parameters according to a neural network architecture using the software and computing resources described above with respect to the data center 800. In at least one embodiment, by using weight parameters calculated by one or more training techniques described herein, information may be inferred or predicted using trained machine learning models corresponding to one or more neural networks using the resources described above with respect to the data center 800.

[0135] In at least one embodiment, the data center can use a CPU, an application-specific integrated circuit (ASIC), a GPU, an FPGA, or other hardware to use the above resources to perform training and / or reasoning. In addition, one or more of the above software and / or hardware resources can be configured as a service to allow users to train or perform information reasoning, such as image recognition, speech recognition, or other artificial intelligence services.

[0136] Reasoning and / or training logic 715 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig. 7A and / or Figure 7BProvides details about the reasoning and / or training logic 715. In at least one embodiment, the reasoning and / or training logic 715 may be implemented in the system Figure 8 for use in a system for performing inference or prediction operations based at least in part on weight parameters computed using neural network training operations, neural network functions and / or architectures, or neural network use cases described herein.

[0137] Such components can be used in animation systems.

[0138] Computer Systems

[0139] Fig. 9 900 is a block diagram illustrating an exemplary computer system according to at least one embodiment, which may be a system of interconnected devices and components, a system on a chip (SOC), or some combination thereof formed with a processor, which may include an execution unit to execute instructions. In at least one embodiment, according to the present disclosure, such as the embodiments described herein, the computer system 900 may include, but is not limited to, components, such as a processor 902, whose execution unit includes logic to execute algorithms for processing data. In at least one embodiment, the computer system 900 may include a processor, such as one available from Intel Corporation of Santa Clara, California. Processor family, Xeon TM , XScale TM and / or Strong ARM TM , Core TM or Nervana TM microprocessor, although other systems (including PCs with other microprocessors, engineering workstations, set-top boxes, etc.) may also be used. In at least one embodiment, computer system 900 may execute a version of the WINDOWS operating system available from Microsoft Corporation of Redmond, Washington, although other operating systems (e.g., UNIX and Linux), embedded software, and / or graphical user interfaces may also be used.

[0140] Embodiments may be used in other devices, such as handheld devices and embedded applications. Some examples of handheld devices include cellular phones, Internet Protocol devices, digital cameras, personal digital assistants ("PDAs"), and handheld PCs. In at least one embodiment, the embedded application may include a microcontroller, a digital signal processor ("DSP"), a system on a chip, a network computer ("NetPC"), a set-top box, a network hub, a wide area network ("WAN") switch, or any other system that can execute one or more instructions according to at least one embodiment.

[0141] In at least one embodiment, the computer system 900 may include, but is not limited to, a processor 902, which may include, but is not limited to, one or more execution units 908 to perform machine learning model training and / or reasoning according to the techniques described herein. In at least one embodiment, the computer system 900 is a single-processor desktop or server system, but in another embodiment, the computer system 900 may be a multi-processor system. In at least one embodiment, the processor 902 may include, but is not limited to, a complex instruction set computing ("CISC") microprocessor, a reduced instruction set computing ("RISC") microprocessor, a very long instruction word ("VLIW") computing microprocessor, a processor that implements an instruction set combination, or any other processor device, such as a digital signal processor. In at least one embodiment, the processor 902 may be coupled to a processor bus 910, which may transmit data signals between the processor 902 and other components in the computer system 900.

[0142] In at least one embodiment, processor 902 may include, but is not limited to, a level 1 ("L1") internal cache memory ("cache") 904. In at least one embodiment, processor 902 may have a single internal cache or multiple levels of internal cache. In at least one embodiment, the cache memory may reside external to processor 902. Other embodiments may also include a combination of internal and external caches, depending on the particular implementation and needs. In at least one embodiment, register file 906 may store different types of data in various registers, including, but not limited to, integer registers, floating point registers, status registers, and instruction pointer registers.

[0143] In at least one embodiment, logic execution unit 908 that performs integer and floating point operations is also located in processor 902, including, but not limited to. In at least one embodiment, processor 902 may also include a microcode ("ucode") read-only memory ("ROM") for storing microcode for certain macroinstructions. In at least one embodiment, execution unit 908 may include logic for processing a packed instruction set 909. In at least one embodiment, by including a packed instruction set 909 in the instruction set of a general purpose processor, and associated circuitry to execute the instructions, operations used by many multimedia applications may be performed using packed data in processor 902. In one or more embodiments, many multimedia applications may be executed faster and more efficiently by using the full width of the processor's data bus to perform operations on packed data, which may not require the transfer of smaller units of data on the processor's data bus to perform one or more operations one data element at a time.

[0144] In at least one embodiment, execution unit 908 may also be used in microcontrollers, embedded processors, graphics devices, DSPs, and other types of logic circuits. In at least one embodiment, computer system 900 may include, but is not limited to, memory 920. In at least one embodiment, memory 920 may be implemented as a dynamic random access memory ("DRAM") device, a static random access memory ("SRAM") device, a flash memory device, or other storage device. In at least one embodiment, memory 920 may store instructions 919 and / or data 921 represented by data signals that may be executed by processor 902.

[0145] In at least one embodiment, the system logic chip can be coupled to the processor bus 910 and the memory 920. In at least one embodiment, the system logic chip can include, but is not limited to, a memory controller hub ("MCH") 916, and the processor 902 can communicate with the MCH 916 via the processor bus 910. In at least one embodiment, the MCH 916 can provide a high-bandwidth memory path 918 to the memory 920 for instruction and data storage and for storage of graphics commands, data, and textures. In at least one embodiment, the MCH 916 can initiate data signals between the processor 902, the memory 920, and other components in the computer system 900, and bridge data signals between the processor bus 910, the memory 920, and the system I / O 922. In at least one embodiment, the system logic chip can provide a graphics port for coupling to a graphics controller. In at least one embodiment, the MCH 916 can be coupled to the memory 920 via a high-bandwidth memory path 918, and the graphics / video card 912 can be coupled to the MCH 916 via an Accelerated Graphics Port ("AGP") interconnect 914.

[0146] In at least one embodiment, the computer system 900 may use a system I / O 922, which is a proprietary hub interface bus to couple the MCH 916 to an I / O controller hub ("ICH") 930. In at least one embodiment, the ICH 930 may provide direct connections to certain I / O devices through a local I / O bus. In at least one embodiment, the local I / O bus may include, but is not limited to, a high-speed I / O bus used to connect peripheral devices to the memory 920, chipset, and processor 902. Examples may include, but are not limited to, an audio controller 929, a firmware hub ("FlashBIOS") 928, a wireless transceiver 926, a data store 924, a traditional I / O controller 923 including a user input and keyboard interface, a serial expansion port 927 (e.g., a universal serial bus (USB) port), and a network controller 934. The data store 924 may include a hard drive, a floppy drive, a CD-ROM device, a flash memory device, or other mass storage device.

[0147] In at least one embodiment, Fig. 9 The system is shown as comprising interconnected hardware devices or "chips", while in other embodiments, Fig. 9 An exemplary system on chip (SoC) may be shown. In at least one embodiment, the devices may be interconnected with a proprietary interconnect, a normalized interconnect (e.g., PCIe), or some combination thereof. In at least one embodiment, one or more components of computer system 900 are interconnected using a compute express link (CXL) interconnect.

[0148] The reasoning and / or training logic 715 is used to perform reasoning and / or training operations related to one or more embodiments. Fig. 7A and / or Figure 7B Provide details about the reasoning and / or training logic 715. In at least one embodiment, the reasoning and / or training logic 715 may be Fig. 9 for use in a system for performing inference or prediction operations based at least in part on weight parameters computed using neural network training operations, neural network functions and / or architectures, or neural network use cases described herein.

[0149] Such components can be used in animation systems.

[0150] Fig.10 1 is a block diagram illustrating an electronic device 1000 for utilizing a processor 1010 according to at least one embodiment. In at least one embodiment, the electronic device 1000 may be, for example but not limited to, a notebook computer, a tower server, a rack server, a blade server, a laptop computer, a desktop computer, a tablet computer, a mobile device, a phone, an embedded computer, or any other suitable electronic device.

[0151] In at least one embodiment, system 1000 may include, but is not limited to, a processor 1010 communicatively coupled to any suitable number or kind of components, peripherals, modules, or devices. In at least one embodiment, processor 1010 is coupled using a bus or interface, such as an I2C bus, a system management bus ("SMBus"), a low pin count (LPC) bus, a serial peripheral interface ("SPI"), a high-definition audio ("HDA") bus, a serial advanced technology attachment ("SATA") bus, a universal serial bus ("USB") (versions 1, 2, 3), or a universal asynchronous receiver / transmitter ("UART") bus. In at least one embodiment, Fig.10 A system is shown that includes interconnected hardware devices or "chips", while in other embodiments, Fig.10 An exemplary system on chip (SoC) may be shown. In at least one embodiment, Fig.10 The devices shown in can be interconnected with a proprietary interconnect, a normalized interconnect (e.g., PCIe), or some combination thereof. In at least one embodiment, Fig.10 One or more components of the system are interconnected using Compute Express Link (CXL) interconnect lines.

[0152] In at least one embodiment, Fig.10 It may include a display 1024, a touch screen 1025, a touch pad 1030, a near field communication unit (“NFC”) 1045, a sensor hub 1040, a thermal sensor 1046, a fast chipset (“EC”) 1035, a trusted platform module (“TPM”) 1038, a BIOS / firmware / flash memory (“BIOS, FWFlash”) 1022, a DSP 1060, a drive 1020 (e.g., a solid state disk (“SSD”) or a hard disk drive (“HDD”)), a wireless local area network unit (“WLAN”) 1050, a Bluetooth unit 1052, a wireless wide area network unit (“WWAN”) 1056, a global positioning system (GPS) 1055, a camera (“USB 3.0 camera”) 1054 (e.g., a USB 3.0 camera), and / or a low power double data rate (“LPDDR”) memory unit (“LPDDR3”) 1015 implemented with, for example, the LPDDR3 standard. These components may each be implemented in any suitable manner.

[0153] In at least one embodiment, other components may be communicatively coupled to the processor 1010 through the components described above. In at least one embodiment, the accelerometer 1041, ambient light sensor (“ALS”) 1042, compass 1043, and gyroscope 1044 may be communicatively coupled to the sensor hub 1040. In at least one embodiment, the thermal sensor 1039, fan 1037, keyboard 1036, and touchpad 1030 may be communicatively coupled to the EC 1035. In at least one embodiment, the speaker 1063, earphone 1064, and microphone (“mic”) 1065 may be communicatively coupled to the audio unit (“audio codec and class D amplifier”) 1062, which in turn may be communicatively coupled to the DSP 1060. In at least one embodiment, the audio unit 1062 may include, for example, but not limited to, an audio encoder / decoder (“codec”) and a class D amplifier. In at least one embodiment, the SIM card (“SIM”) 1057 may be communicatively coupled to the WWAN unit 1056. In at least one embodiment, components such as the WLAN unit 1050 and the Bluetooth unit 1052 and the WWAN unit 1056 may be implemented as a next generation form factor (NGFF).

[0154] The reasoning and / or training logic 715 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig. 7A and / or Figure 7B Provide details about the reasoning and / or training logic 715. In at least one embodiment, the reasoning and / or training logic 715 may be Fig.10 Systems for performing inference or prediction operations based at least in part on weight parameters computed using neural network training operations, neural network functions and / or architectures, or neural network use cases described herein.

[0155] Such components can be used in animation systems.

[0156] Fig.11 1100 is a block diagram of a processing system according to at least one embodiment. In at least one embodiment, system 1100 includes one or more processors 1102 and one or more graphics processors 1108, and can be a single-processor desktop system, a multi-processor workstation system, or a server system with a large number of processors 1102 or processor cores 1107. In at least one embodiment, system 1100 is a processing platform incorporated within a system-on-chip (SoC) integrated circuit for use in a mobile, handheld, or embedded device.

[0157] In at least one embodiment, the system 1100 may include or be incorporated into a server-based gaming platform, including a gaming console, a mobile gaming console, a handheld gaming console, or an online gaming console for gaming and media consoles. In at least one embodiment, the system 1100 is a mobile phone, a smart phone, a tablet computing device, or a mobile Internet device. In at least one embodiment, the processing system 1100 may also include a wearable device coupled to or integrated in a wearable device, such as a smart watch wearable device, a smart glasses device, an augmented reality device, or a virtual reality device. In at least one embodiment, the processing system 1100 is a television or set-top box device having one or more processors 1102 and a graphical interface generated by one or more graphics processors 1108.

[0158] In at least one embodiment, one or more processors 1102 each include one or more processor cores 1107 to process instructions that, when executed, perform operations for system and user software. In at least one embodiment, each of the one or more processor cores 1107 is configured to process a specific instruction group 1109. In at least one embodiment, the instruction group 1109 can facilitate complex instruction set computing (CISC), reduced instruction set computing (RISC), or calculate through very long instruction words (VLIW). In at least one embodiment, the processor cores 1107 can each process different instruction groups 1109, which can include instructions that help simulate other instruction groups. In at least one embodiment, the processor core 1107 can also include other processing devices, such as a digital signal processor (DSP).

[0159] In at least one embodiment, one or more processors 1102 include cache memory 1104. In at least one embodiment, one or more processors 1102 may have a single internal cache or multiple levels of internal cache. In at least one embodiment, the cache memory is shared between various components of the processor 1102. In at least one embodiment, one or more processors 1102 also use an external cache (e.g., a level 3 (L3) cache or a last level cache (LLC)) (not shown), which may be shared between one or more processor cores 1107 using known cache coherence techniques. In at least one embodiment, one or more processors 1102 additionally include a register file 1106, and the processor may include different types of registers (e.g., integer registers, floating point registers, status registers, and instruction pointer registers) for storing different types of data. In at least one embodiment, the register file 1106 may include general registers or other registers.

[0160] In at least one embodiment, one or more processors 1102 are coupled to one or more interface buses 1110 to transmit communication signals, such as address, data, or control signals, between one or more processors 1102 and other components in the system 1100. In at least one embodiment, one or more interface buses 1110 may be a processor bus in one embodiment, such as a version of a direct media interface (DMI) bus. In at least one embodiment, one or more interface buses 1110 are not limited to a DMI bus, and may include one or more peripheral component interconnect buses (e.g., PCI, PCI Express), a memory bus, or other types of interface buses. In at least one embodiment, the processor 1102 includes an integrated memory controller 1116 and a platform controller hub 1130. In at least one embodiment, the memory controller 1116 facilitates communication between memory devices and other components of the processing system 1100, while the platform controller hub (PCH) 1130 provides connections to I / O devices through a local I / O bus.

[0161] In at least one embodiment, the memory device 1120 may be a dynamic random access memory (DRAM) device, a static random access memory (SRAM) device, a flash memory device, a phase change memory device, or have appropriate performance to be used as a processor memory. In at least one embodiment, the memory device 1120 may be used as a system memory of the processing system 1100 to store data 1122 and instructions 1121 for use when one or more processors 1102 execute an application or process. In at least one embodiment, the memory controller 1116 is also coupled to an optional external graphics processor 1112, which may communicate with one or more graphics processors 1108 in the one or more processors 1102 to perform graphics and media operations. In at least one embodiment, a display device 1111 may be connected to the one or more processors 1102. In at least one embodiment, the display device 1111 may include one or more of the internal display devices, such as in a mobile electronic device or laptop device or an external display device connected via a display interface (e.g., DisplayPort, etc.). In at least one embodiment, the display device 1111 may include a head mounted display (HMD), such as a stereoscopic display device used in virtual reality (VR) applications or augmented reality (AR) applications.

[0162] In at least one embodiment, the platform controller hub 1130 enables peripheral devices to be connected to the storage device 1120 and one or more processors 1102 via a high-speed I / O bus. In at least one embodiment, the I / O peripherals include, but are not limited to, an audio controller 1146, a network controller 1134, a firmware interface 1128, a wireless transceiver 1126, a touch sensor 1125, a data storage device 1124 (e.g., a hard drive, flash memory, etc.). In at least one embodiment, the data storage device 1124 can be connected via a storage interface (e.g., SATA) or via a peripheral bus, such as a peripheral component interconnect bus (e.g., PCI, PCIe). In at least one embodiment, the touch sensor 1125 can include a touch screen sensor, a pressure sensor, or a fingerprint sensor. In at least one embodiment, the wireless transceiver 1126 can be a Wi-Fi transceiver, a Bluetooth transceiver, or a mobile network transceiver, such as a 3G, 4G, or long-term evolution (LTE) transceiver. In at least one embodiment, the firmware interface 1128 enables communication with the system firmware and can be, for example, a unified extensible firmware interface (UEFI). In at least one embodiment, the network controller 1134 can enable a network connection to a wired network. In at least one embodiment, a high-performance network controller (not shown) is coupled to one or more interface buses 1110. In at least one embodiment, the audio controller 1146 is a multi-channel high-definition audio controller. In at least one embodiment, the processing system 1100 includes an optional legacy I / O controller 1140 for coupling legacy (e.g., Personal System 2 (PS / 2)) devices to the system 1100. In at least one embodiment, the platform controller hub 1130 can also be connected to one or more universal serial bus (USB) controllers 1142, which connect input devices such as a keyboard and mouse 1143 combination, a camera 1144, or other USB input devices.

[0163] In at least one embodiment, instances of memory controller 1116 and platform controller hub 1130 may be integrated into a discrete external graphics processor, such as external graphics processor 1112. In at least one embodiment, platform controller hub 1130 and / or memory controller 1116 may be external to one or more processors 1102. For example, in at least one embodiment, system 1100 may include external memory controller 1116 and platform controller hub 1130, which may be configured as a memory controller hub and a peripheral controller hub in a system chipset that communicates with processor 1102.

[0164] The reasoning and / or training logic 715 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig. 7A and / or Figure 7B Detail is provided regarding the inference and / or training logic 715. In at least one embodiment, some or all of the inference and / or training logic 715 may be incorporated into the graphics processor 1100. For example, in at least one embodiment, the training and / or inference techniques described herein may use one or more ALUs that are embodied in the graphics processor. In addition, in at least one embodiment, the inference and / or training operations described herein may use the ALUs other than the ALUs. Fig. 7A and / or Figure 7B In at least one embodiment, the weight parameters may be stored in on-chip or off-chip memory and / or registers (shown or not shown) that configure the ALU of the graphics processor to execute one or more of the machine learning algorithms, neural network architectures, use cases, or training techniques described herein.

[0165] Such components can be used in animation systems.

[0166] Fig.12 is a block diagram of a processor 1200 having one or more processor cores 1202A-1202N, an integrated memory controller 1214, and an integrated graphics processor 1208, according to at least one embodiment. In at least one embodiment, the processor 1200 may include additional cores, up to and including the additional core 1202N represented by the dashed box. In at least one embodiment, each processor core 1202A-1202N includes one or more internal cache units 1204A-1204N. In at least one embodiment, each processor core may also have access to one or more shared cache units 1206.

[0167] In at least one embodiment, one or more internal cache units 1204A-1204N and one or more shared cache units 1206 represent a cache memory hierarchy within the processor 1200. In at least one embodiment, the one or more cache units 1204A-1204N may include at least one level of instruction and data cache within each processor core and one or more levels of cache in a shared mid-level cache, such as level 2 (L2), level 3 (L3), level 4 (L4), or other levels of cache, where the highest level of cache before external memory is classified as LLC. In at least one embodiment, cache coherency logic maintains coherency between the various cache units 1206 and 1204A-1204N.

[0168] In at least one embodiment, the processor 1200 may also include a set of one or more bus controller units 1216 and a system agent core 1210. In at least one embodiment, the one or more bus controller units 1216 manage a set of peripheral buses, such as one or more PCI or PCIe buses. In at least one embodiment, the system agent core 1210 provides management functions for various processor components. In at least one embodiment, the system agent core 1210 includes one or more integrated memory controllers 1214 to manage access to various external memory devices (not shown).

[0169] In at least one embodiment, one or more processor cores 1202A-1202N include support for multiple threads simultaneously. In at least one embodiment, system agent core 1210 includes components for coordination during multithreaded processing and one or more processor cores 1202A-1202N. In at least one embodiment, system agent core 1210 may additionally include a power control unit (PCU) including logic and components for regulating one or more power states of one or more processor cores 1202A-1202N and graphics processor 1208.

[0170] In at least one embodiment, the processor 1200 also includes a graphics processor 1208 for performing graphics processing operations. In at least one embodiment, the graphics processor 1208 is coupled to one or more shared cache units 1206 and a system agent core 1210 including one or more integrated memory controllers 1214. In at least one embodiment, the system agent core 1210 also includes a display controller 1211 for driving the graphics processor output to one or more coupled displays. In at least one embodiment, the display controller 1211 may also be a separate module coupled to the graphics processor 1208 via at least one interconnect, or may be integrated within the graphics processor 1208.

[0171] In at least one embodiment, a ring-based interconnect unit 1212 is used to couple the internal components of the processor 1200. In at least one embodiment, alternative interconnect units may be used, such as point-to-point interconnects, switched interconnects, or other technologies. In at least one embodiment, the graphics processor 1208 is coupled to the ring-based interconnect unit 1212 via an I / O link 1213.

[0172] In at least one embodiment, I / O link 1213 represents at least one of a variety of I / O interconnects, including packaged I / O interconnects that facilitate communication between various processor components and high-performance embedded memory modules 1218 (e.g., eDRAM modules). In at least one embodiment, each of one or more processor cores 1202A-1202N and graphics processor 1208 uses embedded memory module 1218 as a shared last level cache.

[0173] In at least one embodiment, one or more processor cores 1202A-1202N are homogeneous cores that execute a common instruction set architecture. In at least one embodiment, one or more processor cores 1202A-1202N are heterogeneous in terms of instruction set architecture (ISA), wherein one or more processor cores 1202A-1202N execute a common instruction set, while one or more other processor cores 1202A-1202N execute a subset or a different instruction set of the common instruction set. In at least one embodiment, one or more processor cores 1202A-1202N are heterogeneous in terms of microarchitecture, wherein one or more cores with relatively high power consumption are coupled with one or more power cores with lower power consumption. In at least one embodiment, processor 1200 can be implemented on one or more chips or implemented as a SoC integrated circuit.

[0174] The reasoning and / or training logic 715 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig. 7A and / or Figure 7B Detail is provided regarding the inference and / or training logic 715. In at least one embodiment, some or all of the inference and / or training logic 715 may be incorporated into the processor 1200. For example, in at least one embodiment, the training and / or inference techniques described herein may use one or more ALUs embodied in Fig.12 In addition, in at least one embodiment, the inference and / or training operations described herein may use the inference and / or training operations described herein. Fig. 7A and / or Figure 7B In at least one embodiment, the weight parameters may be stored in on-chip or off-chip memory and / or registers (shown or not shown) that configure the ALU of the graphics processor 1200 to execute one or more of the machine learning algorithms, neural network architectures, use cases, or training techniques described herein.

[0175] Such components can be used in animation systems.

[0176] Virtualized computing platform

[0177] Fig.13 is an example data flow diagram of a process 1300 for generating and deploying an image processing and inference pipeline according to at least one embodiment. In at least one embodiment, the process 1300 can be deployed for use with imaging devices, processing devices, and / or other device types at one or more facilities 1302. The process 1300 can be executed within a training system 1304 and / or a deployment system 1306. In at least one embodiment, the training system 1304 can be used to perform training, deployment, and implementation of machine learning models (e.g., neural networks, object detection algorithms, computer vision algorithms, etc.) for use with the deployment system 1306. In at least one embodiment, the deployment system 1306 can be configured to offload processing and computing resources in a distributed computing environment to reduce infrastructure requirements of the facility 1302. In at least one embodiment, one or more applications in the pipeline can use or call services (e.g., reasoning, visualization, computation, AI, etc.) of the deployment system 1306 during application execution.

[0178] In at least one embodiment, some applications used in the high-level processing and reasoning pipeline may use machine learning models or other AI to perform one or more processing steps. In at least one embodiment, the machine learning model may be trained at the facility 1302 using data 1308 (e.g., imaging data) generated at the facility 1302 (and stored on one or more picture archiving and communication system (PACS) servers at the facility 1302), the machine learning model may be trained using imaging or sequencing data 1308 from another one or more facilities, or a combination thereof. In at least one embodiment, the training system 1304 may be used to provide applications, services, and / or other resources to generate a working, deployable machine learning model for the deployment system 1306.

[0179] In at least one embodiment, the model registry 1324 can be backed by an object store, which can support versioning and object metadata. In at least one embodiment, the object store can be accessed from within the cloud platform through, for example, a cloud storage compatible application programming interface (API). In at least one embodiment, the machine learning models within the model registry 1324 can be uploaded, listed, modified, or deleted by developers or partners of the system interacting with the API. In at least one embodiment, the API can provide access to methods that allow users with appropriate credentials to associate a model with an application so that the model can be executed as part of the execution of a containerized instantiation of the application.

[0180] In at least one embodiment, the training system 1304 ( Fig.13) may include situations where the facility 1302 is training their own machine learning model, or has an existing machine learning model that needs to be optimized or updated. In at least one embodiment, imaging data 1308 generated by an imaging device, a sequencing device, and / or other type of device may be received. In at least one embodiment, once the imaging data 1308 is received, the AI-assisted annotation 1310 may be used to help generate annotations corresponding to the imaging data 1308 to be used as ground truth data for the machine learning model. In at least one embodiment, the AI-assisted annotation 1310 may include one or more machine learning models (e.g., a convolutional neural network (CNN)), which may be trained to generate annotations corresponding to certain types of imaging data 1308 (e.g., from certain devices). In at least one embodiment, the AI-assisted annotation 1310 may then be used directly, or may be adjusted or fine-tuned using an annotation tool to generate ground truth data. In at least one embodiment, the AI-assisted annotation 1310, the labeled data 1312, or a combination thereof may be used as ground truth data for training a machine learning model. In at least one embodiment, the trained machine learning models may be referred to as one or more output models 1316 and may be used by the deployment system 1306 as described herein.

[0181] In at least one embodiment, the training pipeline may include a scenario where the facility 1302 requires a machine learning model for performing one or more processing tasks for one or more applications in the deployment system 1306, but the facility 1302 may not currently have such a machine learning model (or may not have a model that is optimized, efficient, or effective for this purpose). In at least one embodiment, an existing machine learning model may be selected from the model registry 1324. In at least one embodiment, the model registry 1324 may include machine learning models that are trained to perform a variety of different reasoning tasks on imaging data. In at least one embodiment, the machine learning models in the model registry 1324 may be trained on imaging data from a different facility (e.g., a facility located far away) rather than the facility 1302. In at least one embodiment, the machine learning model may have been trained on imaging data from one location, two locations, or any number of locations. In at least one embodiment, when training is performed on imaging data from a specific location, the training may be performed at that location, or at least in a manner that protects the confidentiality of the imaging data or restricts the transfer of the imaging data from off-site. In at least one embodiment, once a model is trained or partially trained at one location, the machine learning model can be added to the model registry 1324. In at least one embodiment, the machine learning model can then be retrained or updated at any number of other facilities, and the retrained or updated model can be used in the model registry 1324. In at least one embodiment, the machine learning model can then be selected from the model registry 1324 (and referred to as one or more output models 1316) and can be deployed in the deployment system 1306 to perform one or more processing tasks for one or more applications of the deployment system.

[0182] In at least one embodiment, the scenario may include a facility 1302 that requires a machine learning model for performing one or more processing tasks for one or more applications in the deployment system 1306, but the facility 1302 may not currently have such a machine learning model (or may not have an optimized, efficient, or effective model). In at least one embodiment, the machine learning model selected from the model registry 1324 may not be fine-tuned or optimized for the imaging data 1308 generated at the facility 1302 due to population differences, robustness of the training data used to train the machine learning model, diversity of training data anomalies, and / or other issues with the training data. In at least one embodiment, AI-assisted annotations 1310 may be used to help generate annotations corresponding to the imaging data 1308 to be used as ground truth data for training or updating the machine learning model. In at least one embodiment, labeled data 1312 may be used as ground truth data for training the machine learning model. In at least one embodiment, retraining or updating the machine learning model may be referred to as model training 1314. In at least one embodiment, model training 1314 (e.g., AI-assisted annotations 1310, labeled data 1312, or a combination thereof) may be used as ground truth data for retraining or updating the machine learning model. In at least one embodiment, the trained machine learning models may be referred to as one or more output models 1316 and may be used by the deployment system 1306 as described herein.

[0183] In at least one embodiment, the deployment system 1306 may include software 1318, services 1320, hardware 1322, and / or other components, features, and functions. In at least one embodiment, the deployment system 1306 may include a software "stack" such that the software 1318 may be built on top of the services 1320 and may use the services 1320 to perform some or all of the processing tasks, and the services 1320 and software 1318 may be built on top of the hardware 1322 and use the hardware 1322 to perform the processing, storage, and / or other computing tasks of the deployment system. In at least one embodiment, the software 1318 may include any number of different containers, each of which may perform an instantiation of an application. In at least one embodiment, each application may perform one or more processing tasks (e.g., reasoning, object detection, feature detection, segmentation, image enhancement, calibration, etc.) in a high-level processing and reasoning pipeline. In at least one embodiment, in addition to receiving and configuring imaging data for use by each container and / or containers used by facility 1302 after processing through the pipeline, a high-level processing and reasoning pipeline can also be defined based on the selection of different containers desired or required to process imaging data 1308 (e.g., to convert the output back into a usable data type. In at least one embodiment, the combination of containers within software 1318 (e.g., which constitute a pipeline) can be referred to as a virtual instrument (as described in more detail herein), and the virtual instrument can utilize services 1320 and hardware 1322 to perform some or all of the processing tasks of an application instantiated in a container.

[0184] In at least one embodiment, the data processing pipeline can receive input data (e.g., imaging data 1308) in a specific format in response to an inference request (e.g., a request from a user of the deployment system 1306). In at least one embodiment, the input data can represent one or more images, videos, and / or other data representations generated by one or more imaging devices. In at least one embodiment, the data can be pre-processed as part of the data processing pipeline to prepare the data for processing by one or more applications. In at least one embodiment, post-processing can be performed on the output of one or more inference tasks or other processing tasks of the pipeline to prepare output data for the next application and / or prepare the output data for transmission and / or use by the user (e.g., as a response to the inference request). In at least one embodiment, the inference task can be performed by one or more machine learning models, such as trained or deployed neural networks, which can include one or more output models 1316 of the training system 1304.

[0185] In at least one embodiment, tasks of a data processing pipeline may be encapsulated in containers, each container representing a discrete, fully functional instantiation of an application and a virtualized computing environment capable of referencing a machine learning model. In at least one embodiment, containers or applications may be published to a private (e.g., limited access) area of ​​a container registry (described in more detail herein), and trained or deployed models may be stored in the model registry 1324 and associated with one or more applications. In at least one embodiment, an image of an application (e.g., a container image) may be available in a container registry, and once a user selects an image from a container registry for deployment in a pipeline, the image may be used to generate an instantiated container for the application for use by the user's system.

[0186] In at least one embodiment, a developer (e.g., a software developer, a clinician, a physician, etc.) can develop, publish, and store applications (e.g., as containers) for performing image processing and / or reasoning on provided data. In at least one embodiment, the development, publishing, and / or storage can be performed using a software development kit (SDK) associated with the system (e.g., to ensure that the developed applications and / or containers conform to or are compatible with the system). In at least one embodiment, the developed applications can be tested locally (e.g., at a first facility, on data from the first facility) using the SDK, which is installed as part of the system (e.g., Fig.14 The system 1400 in FIG. 14 may support at least some of the services 1320. In at least one embodiment, because a DICOM object may contain from one to hundreds of images or other data types, and because the data varies, the developer may be responsible for managing (e.g., setting up constructs for building pre-processing into the application, etc.) the extraction and preparation of the incoming data. In at least one embodiment, once validated by the system 1400 (e.g., for accuracy), the application is made available in the container registry for selection and / or implementation by the user to perform one or more processing tasks on the data at the user's facility (e.g., a second facility).

[0187] In at least one embodiment, the developer can then share the application or container over a network for use by a system (e.g., Fig.141400). In at least one embodiment, the completed and validated application or container can be stored in the container registry, and the associated machine learning model can be stored in the model registry 1324. In at least one embodiment, the requesting entity (which provides an inference or image processing request) can browse the container registry and / or the model registry 1324 to obtain applications, containers, data sets, machine learning models, etc., select the desired combination of elements to be included in the data processing pipeline, and submit the image processing request. In at least one embodiment, the request may include input data necessary to execute the request (and patient-related data in some examples), and / or may include a selection of applications and / or machine learning models to be executed when processing the request. In at least one embodiment, the request can then be passed to one or more components (e.g., a cloud) of the deployment system 1306 to perform processing of the data processing pipeline. In at least one embodiment, the processing performed by the deployment system 1306 may include referencing elements (e.g., applications, containers, models, etc.) selected from the container registry and / or the model registry 1324. In at least one embodiment, once the results are generated by the pipeline, the results may be returned to the user for reference (eg, for viewing in a viewing application suite executing locally, on a local workstation or terminal).

[0188] In at least one embodiment, to assist in processing or executing applications or containers in the pipeline, services 1320 may be utilized. In at least one embodiment, services 1320 may include computing services, artificial intelligence (AI) services, visualization services, and / or other service types. In at least one embodiment, services 1320 may provide functionality common to one or more applications in software 1318, and thus may abstract functionality into services that may be called or utilized by applications. In at least one embodiment, the functionality provided by services 1320 may run dynamically and more efficiently, while also allowing applications to process data in parallel (e.g., using Fig.14In at least one embodiment, rather than requiring that each application that shares the same functionality provided by the service 1320 must have a corresponding instance of the service 1320, the service 1320 can be shared between and among various applications. In at least one embodiment, as a non-limiting example, the service may include an inference server or engine that can be used to perform detection or segmentation tasks. In at least one embodiment, a model training service may be included that can provide machine learning model training and / or retraining capabilities. In at least one embodiment, a data enhancement service may be further included that can provide GPU-accelerated data (e.g., DICOM, RIS, CIS, REST-compliant, RPC, raw, etc.) extraction, resizing, scaling, and / or other enhancements. In at least one embodiment, a visualization service may be used that can add image rendering effects (e.g., ray tracing, rasterization, denoising, sharpening, etc.) to add realism to two-dimensional (2D) and / or three-dimensional (3D) models. In at least one embodiment, a virtual instrument service may be included that provides beamforming, segmentation, reasoning, imaging, and / or support for other applications within the pipeline of the virtual instrument.

[0189] In at least one embodiment, where the service 1320 includes an AI service (e.g., an inference service), as part of the execution of an application, one or more machine learning models can be executed by calling (e.g., as an API call) an inference service (e.g., an inference server) to execute one or more machine learning models or processing thereof. In at least one embodiment, where another application includes one or more machine learning models for a segmentation task, the application can call the inference service to execute the machine learning model for performing one or more processing operations associated with the segmentation task. In at least one embodiment, the software 1318 implementing the high-level processing and inference pipeline, which includes a segmentation application and anomaly detection application, can be pipelined in that each application can call the same inference service to perform one or more inference tasks.

[0190] In at least one embodiment, the hardware 1322 may include a GPU, a CPU, a graphics card, an AI / deep learning system (e.g., an AI supercomputer such as NVIDIA's DGX), a cloud platform, or a combination thereof. In at least one embodiment, different types of hardware 1322 may be used to provide efficient, purpose-built support for the software 1318 and services 1320 in the deployment system 1306. In at least one embodiment, the use of GPU processing may be implemented to perform local processing (e.g., at the facility 1302) within the AI / deep learning system, in the cloud system, and / or other processing components of the deployment system 1306 to improve the efficiency, accuracy, and effectiveness of image processing and generation. In at least one embodiment, as a non-limiting example, with respect to deep learning, machine learning, and / or high-performance computing, the software 1318 and / or services 1320 may be optimized for GPU processing. In at least one embodiment, at least some of the computing environments of the deployment system 1306 and / or training system 1304 may be executed in a data center, one or more supercomputers, or high-performance computer systems with GPU-optimized software (e.g., a combination of hardware and software of an NVIDIA DGX system). In at least one embodiment, as described herein, hardware 1322 may include any number of GPUs that may be called to perform data processing in parallel. In at least one embodiment, the cloud platform may also include GPU optimized execution for deep learning tasks, GPU processing for machine learning tasks or other computing tasks. In at least one embodiment, an AI / deep learning supercomputer and / or GPU optimized software (e.g., as provided on NVIDIA's DGX system) may be used as a hardware abstraction and scaling platform to execute a cloud platform (e.g., NVIDIA's NGC). In at least one embodiment, the cloud platform may integrate an application container cluster system or coordination system (e.g., KUBERNETES) on multiple GPUs to achieve seamless scaling and load balancing.

[0191] Fig.14 is a system diagram of an example system 1400 for generating and deploying an imaging deployment pipeline according to at least one embodiment. In at least one embodiment, the system 1400 can be used to implement Fig.13 The process 1300 and / or other processes of the system 1400 may include a high-level processing and reasoning pipeline. In at least one embodiment, the system 1400 may include a training system 1304 and a deployment system 1306. In at least one embodiment, the training system 1304 and the deployment system 1306 may be implemented using software 1318, services 1320, and / or hardware 1322, as described herein.

[0192] In at least one embodiment, system 1400 (e.g., training system 1304 and / or deployment system 1306) can be implemented in a cloud computing environment (e.g., using cloud 1426). In at least one embodiment, system 1400 can be implemented locally (with respect to a healthcare service facility) or as a combination of cloud computing resources and local computing resources. In at least one embodiment, access to APIs in cloud 1426 can be restricted to authorized users by establishing security measures or protocols. In at least one embodiment, the security protocol can include a network token, which can be signed by an authentication (e.g., AuthN, AuthZ, Gluecon, etc.) service and can carry appropriate authorization. In at least one embodiment, the API of the virtual instrument (described herein) or other instances of system 1400 can be restricted to a set of public IPs that have been audited or authorized for interaction.

[0193] In at least one embodiment, the various components of system 1400 can communicate with each other using any of a variety of different network types, including but not limited to local area networks (LANs) and / or wide area networks (WANs) via wired and / or wireless communication protocols. In at least one embodiment, communications between facilities and components of system 1400 (e.g., for sending inference requests, for receiving results of inference requests, etc.) can be transmitted via one or more data buses, wireless data protocols (Wi-Fi), wired data protocols (e.g., Ethernet), etc.

[0194] In at least one embodiment, similar to the present disclosure regarding Fig.13 As described, the training system 1304 can execute the training pipeline 1404. In at least one embodiment, where the deployment system 1306 will use one or more machine learning models in one or more deployment pipelines 1410, the training pipeline 1404 can be used to train or retrain one or more (e.g., pre-trained) models, and / or implement one or more pre-trained models 1406 (e.g., without retraining or updating). In at least one embodiment, as a result of the training pipeline 1404, an output model 1316 can be generated. In at least one embodiment, the training pipeline 1404 can include any number of processing steps, such as, but not limited to, conversion or adaptation of imaging data (or other input data). In at least one embodiment, different training pipelines 1404 can be used for different machine learning models used by the deployment system 1306. In at least one embodiment, similar to the description regarding Fig.13 The first example training pipeline 1404 described may be used for a first machine learning model, similar to the training pipeline 1404 described with respect to FIG. Fig.13 The second example training pipeline 1404 described may be used for a second machine learning model, similar to the one described with respect to Fig.13The training pipeline 1404 of the third example described may be used for a third machine learning model. In at least one embodiment, any combination of tasks within the training system 1304 may be used according to the requirements of each corresponding machine learning model. In at least one embodiment, one or more machine learning models may have been trained and ready for deployment, so the training system 1304 may not perform any processing on the machine learning model, and one or more machine learning models may be implemented by the deployment system 1306.

[0195] In at least one embodiment, the output model 1316 and / or the pre-trained model 1406 may include any type of machine learning model, depending on the implementation or embodiment. In at least one embodiment and without limitation, the machine learning model used by the system 1400 may include using linear regression, logistic regression, decision tree, support vector machine (SVM), naive Bayes, k-nearest neighbor (Knn), k-means clustering, random forest, dimensionality reduction algorithm, gradient boosting algorithm, neural network (e.g., autoencoder, convolution, recursion, perceptron, long / short term memory (LSTM), Hopfield, Boltzmann, deep belief, deconvolution, generative adversarial, liquid state machine, etc.), and / or other types of machine learning models.

[0196] In at least one embodiment, the training pipeline 1404 may include AI-assisted annotation, as described herein with respect to at least Fig. 15BIn at least one embodiment, the labeled data 1312 (e.g., traditional annotations) may be generated by any number of techniques. In at least one embodiment, in some examples, labels or other annotations may be generated in a drawing program (e.g., an annotation program), a computer-aided design (CAD) program, a labeling program, another type of application suitable for generating annotations or labels for ground truth, and / or may be hand-drawn. In at least one embodiment, ground truth data may be synthetically generated (e.g., generated from a computer model or rendering), truly generated (e.g., designed and generated from real-world data), machine-generated (e.g., extracting features from data using feature analysis and learning, and then generating labels), manually annotated (e.g., a labeler or annotation expert, defining the location of the label), and / or a combination thereof. In at least one embodiment, for each instance of imaging data 1308 (or other data types used by machine learning models), there may be corresponding ground truth data generated by the training system 1304. In at least one embodiment, AI-assisted annotation may be performed as part of one or more deployment pipelines 1410; supplementing or replacing the AI-assisted annotation included in the training pipeline 1404. In at least one embodiment, the system 1400 may include a multi-layer platform that may include a software layer (e.g., software 1318) of a diagnostic application (or other application type) that may perform one or more medical imaging and diagnostic functions. In at least one embodiment, the system 1400 may be communicatively coupled (e.g., via an encrypted link) to a PACS server network of one or more facilities. In at least one embodiment, the system 1400 may be configured to access and reference data from a PACS server to perform operations such as training a machine learning model, deploying a machine learning model, image processing, reasoning, and / or other operations.

[0197] In at least one embodiment, the software layer may be implemented as a secure, encrypted, and / or authenticated API through which an application or container may be invoked (e.g., called) from an external environment (e.g., facility 1302). In at least one embodiment, the application may then call or execute one or more services 1320 to perform computational, AI, or visualization tasks associated with the respective application, and the software 1318 and / or services 1320 may utilize hardware 1322 to perform processing tasks in an effective and efficient manner. In at least one embodiment, communications sent to or received by the training system 1304 and the deployment system 1306 may occur using a pair of DICOM adapters 1402A, 1402B.

[0198] In at least one embodiment, the deployment system 1306 can execute a deployment pipeline 1410. In at least one embodiment, the deployment pipeline 1410 can include any number of applications, which can be sequential, non-sequential, or otherwise applied to imaging data (and / or other data types) - including AI-assisted annotations, generated by imaging devices, sequencing devices, genomics devices, etc., as described above. In at least one embodiment, as described herein, one or more deployment pipelines 1410 for individual devices can be referred to as virtual instruments for the device (e.g., virtual ultrasound instruments, virtual CT scanning instruments, virtual sequencing instruments, etc.). In at least one embodiment, there can be more than one deployment pipeline 1410 for a single device, depending on the information desired from the data generated by the device. In at least one embodiment, there can be a first deployment pipeline 1410 in the case where an abnormality is expected to be detected from an MRI machine, and there can be a second deployment pipeline 1410 in the case where image enhancement is expected from the output of the MRI machine.

[0199] In at least one embodiment, the image generation application may include a processing task that includes the use of a machine learning model. In at least one embodiment, a user may wish to use their own machine learning model, or select a machine learning model from the model registry 1324. In at least one embodiment, a user may implement their own machine learning model or select a machine learning model to be included in an application that performs a processing task. In at least one embodiment, the application may be selectable and customizable, and by defining the construction of the application, the deployment and implementation of the application for a particular user is presented as a more seamless user experience. In at least one embodiment, by leveraging other features of the system 1400 (e.g., services 1320 and hardware 1322), one or more deployment pipelines 1410 may be more user-friendly, provide easier integration, and produce more accurate, efficient, and timely results.

[0200] In at least one embodiment, the deployment system 1306 may include a user interface (UI) 1414 (e.g., a graphical user interface, a web interface, etc.) that may be used to select applications to be included in the deployment pipeline 1410, to arrange applications, to modify or change applications or their parameters or configuration, to use and interact with the deployment pipeline 1410 during setup and / or deployment, and / or to otherwise interact with the deployment system 1306. In at least one embodiment, although not shown with respect to the training system 1304, the UI 1414 (or a different user interface) may be used to select models for use in the deployment system 1306, to select models for training or retraining in the training system 1304, and / or to otherwise interact with the training system 1304.

[0201] In at least one embodiment, in addition to the application coordination system 1428, a pipeline manager 1412 may be used to manage the interaction between the application or container of the deployment pipeline 1410 and the service 1320 and / or the hardware 1322. In at least one embodiment, the pipeline manager 1412 may be configured to facilitate the interaction from application to application, from application to service 1320, and / or from application or service to hardware 1322. In at least one embodiment, although shown as included in the software 1318, this is not intended to be limiting, and in some examples, the pipeline manager 1412 may be included in the service 1320. In at least one embodiment, the application coordination system 1428 (e.g., Kubernetes, DOCKER, etc.) may include a container coordination system that can group applications into containers as a logical unit for coordination, management, scaling, and deployment. In at least one embodiment, by associating applications from the deployment pipeline 1410 (e.g., rebuilding applications, splitting applications, etc.) with individual containers, each application can be executed in a self-contained environment (e.g., at the kernel level) to improve speed and efficiency.

[0202] In at least one embodiment, each application and / or container (or its image) can be developed, modified, and deployed separately (e.g., a first user or developer can develop, modify, and deploy a first application, and a second user or developer can develop, modify, and deploy a second application separate from the first user or developer), which can allow for focus and attention on the tasks of a single application and / or container without being hindered by the tasks of another application or container. In at least one embodiment, the pipeline manager 1412 and the application coordination system 1428 can assist in communication and collaboration between different containers or applications. In at least one embodiment, as long as the expected inputs and / or outputs of each container or application are known to the system (e.g., based on the construction of the application or container), the application coordination system 1428 and / or the pipeline manager 1412 can facilitate communication and sharing of resources between and among each application or container. In at least one embodiment, since one or more applications or containers in the deployment pipeline 1410 can share the same services and resources, the application coordination system 1428 can coordinate, load balance, and determine the sharing of services or resources between and among the various applications or containers. In at least one embodiment, the scheduler can be used to track resource requirements of applications or containers, current or planned use of those resources, and resource availability. Thus, in at least one embodiment, the scheduler can allocate resources to different applications and allocate resources between and among applications, taking into account the needs and availability of the system. In some examples, the scheduler (and / or other components of the application coordination system 1428) can determine resource availability and distribution based on constraints imposed on the system (e.g., user constraints), such as quality of service (QoS), the urgency of data output (e.g., to determine whether to perform real-time processing or delayed processing), etc.

[0203] In at least one embodiment, the services 1320 utilized and shared by the applications or containers in the deployment system 1306 may include one or more computing services 1416, one or more AI services 1418, one or more visualization services 1420, and / or other service types. In at least one embodiment, the application may call (e.g., execute) one or more services 1320 to perform processing operations for the application. In at least one embodiment, the application may utilize one or more computing services 1416 to perform supercomputing or other high performance computing (HPC) tasks. In at least one embodiment, one or more computing services 1416 may be utilized to perform parallel processing (e.g., using a parallel computing platform 1430) to process data substantially simultaneously by one or more applications and / or one or more tasks of a single application. In at least one embodiment, a parallel computing platform 1430 (e.g., NVIDIA's CUDA) may implement general computing on a GPU (GPGPU) (e.g., GPU / graphics 1422). In at least one embodiment, the software layer of the parallel computing platform 1430 may provide access to a virtual instruction set and parallel computing elements of a GPU to execute computing kernels. In at least one embodiment, the parallel computing platform 1430 may include memory, and in some embodiments, memory may be shared between and among multiple containers, and / or between and among different processing tasks within a single container. In at least one embodiment, inter-process communication (IPC) calls may be generated for multiple containers and / or multiple processes within a container to enable the same data (e.g., where multiple different stages of an application or multiple applications are processing the same information) to be used for a shared memory segment from the parallel computing platform 1430. In at least one embodiment, instead of copying data and moving the data to different locations in memory (e.g., read / write operations), the same data in the same location of the memory may be used for any number of processing tasks (e.g., at the same time, at different times, etc.). In at least one embodiment, since the data as a result of processing is used to generate new data, this information of the new location of the data may be stored and shared between various applications. In at least one embodiment, the location of the data and the location of the updated or modified data may be part of the definition of how to understand the payload in the container.

[0204] In at least one embodiment, one or more AI services 1418 may be utilized to perform reasoning services for executing machine learning models associated with an application (e.g., the task is to execute one or more processing tasks of the application). In at least one embodiment, one or more AI services 1418 may utilize an AI system 1424 to execute machine learning models (e.g., neural networks such as CNN) for segmentation, reconstruction, object detection, feature detection, classification, and / or other reasoning tasks. In at least one embodiment, an application of the deployment pipeline 1410 may use one or more output models 1316 and / or other models of the application for the self-training system 1304 to perform reasoning on imaging data. In at least one embodiment, two or more examples of reasoning using an application coordination system 1428 (e.g., a scheduler) may be available. In at least one embodiment, a first category may include a high priority / low latency path that can achieve a higher service level agreement, such as for performing reasoning on an urgent request in an emergency, or for a radiologist during a diagnostic process. In at least one embodiment, a second category may include a standard priority path that may be used for requests that may not be urgent or for situations where analysis can be performed at a later time. In at least one embodiment, application coordination system 1428 may allocate resources (e.g., services 1320 and / or hardware 1322) for different reasoning tasks of one or more AI services 1418 based on priority paths.

[0205] In at least one embodiment, the shared memory may be mounted to one or more AI services 1418 in the system 1400. In at least one embodiment, the shared memory may operate as a cache (or other storage device type) and may be used to process inference requests from applications. In at least one embodiment, when an inference request is submitted, a set of API instances of the deployment system 1306 may receive the request and may select one or more instances (e.g., for best fit, for load balancing, etc.) to process the request. In at least one embodiment, in order to process the request, the request may be entered into a database, and if it is not already in the cache, the machine learning model may be located from the model registry 1324, a verification step may ensure that the appropriate machine learning model is loaded into the cache (e.g., shared storage), and / or a copy of the model may be saved to the cache. In at least one embodiment, if the application is not yet running or there are not enough instances of the application, a scheduler (e.g., a scheduler of the pipeline manager 1412) may be used to start the application referenced in the request. In at least one embodiment, if the inference server has not yet been started to execute the model, the inference server may be started. Any number of inference servers may be started for each model. In at least one embodiment, in a pull model that clusters inference servers, the model can be cached whenever load balancing is beneficial. In at least one embodiment, the inference servers can be statically loaded into the corresponding distributed servers.

[0206] In at least one embodiment, inference can be performed using an inference server running in a container. In at least one embodiment, an instance of an inference server can be associated with a model (and optionally with multiple versions of a model). In at least one embodiment, if an instance of an inference server does not exist when a request is received to perform inference on a model, a new instance can be loaded. In at least one embodiment, when the inference server is started, the model can be passed to the inference server so that the same container can be used to serve different models as long as the inference server is running as a different instance.

[0207] In at least one embodiment, during application execution, a request for reasoning for a given application may be received, and a container (e.g., an instance of a hosted reasoning server) may be loaded (if not already loaded), and a launcher may be called. In at least one embodiment, preprocessing logic in the container may load, decode, and / or perform any additional preprocessing on the incoming data (e.g., using a CPU and / or GPU). In at least one embodiment, once the data is ready for reasoning, the container may perform reasoning on the data as needed. In at least one embodiment, this may include a single reasoning call for one image (e.g., a hand X-ray), or may require reasoning for hundreds of images (e.g., a chest CT). In at least one embodiment, the application may summarize the results before completion, which may include, but is not limited to, a single confidence score, pixel-level segmentation, voxel-level segmentation, generating visualizations, or generating text to summarize the results. In at least one embodiment, different priorities may be assigned to different models or applications. For example, some models may have a real-time (TAT less than 1 minute) priority, while other models may have a lower priority (e.g., TAT less than 10 minutes). In at least one embodiment, model execution time may be measured from a requesting mechanism or entity and may include collaborative network traversal time as well as execution time of an inference service.

[0208] In at least one embodiment, the transmission of requests between the service 1320 and the reasoning application can be hidden behind a software development kit (SDK), and robust transmission can be provided through queues. In at least one embodiment, requests will be placed in queues through an API for individual application / tenant ID combinations, and the SDK will pull requests from the queue and provide the request to the application. In at least one embodiment, the name of the queue can be provided in the environment from which the SDK will pick up the queue. In at least one embodiment, asynchronous communication through queues may be useful because it can allow any instance of the application to pick up work when it is available. The results can be transmitted back through the queue to ensure that there is no data loss. In at least one embodiment, the queue can also provide the ability to split the work, because the highest priority work can enter the queue connected to most instances of the application, and the lowest priority work can enter the queue connected to a single instance, which processes the tasks in the order received. In at least one embodiment, the application can run on a GPU-accelerated instance generated in the cloud 1426, and the reasoning service can perform reasoning on the GPU.

[0209] In at least one embodiment, one or more visualization services 1420 may be utilized to generate visualizations for viewing application and / or deployment pipeline 1410 outputs. In at least one embodiment, one or more visualization services 1420 may utilize GPU / graphics 1422 to generate visualizations. In at least one embodiment, one or more visualization services 1420 may implement rendering effects such as ray tracing to generate higher quality visualizations. In at least one embodiment, visualizations may include, but are not limited to, 2D image rendering, 3D volume rendering, 3D volume reconstruction, 2D tomographic slices, virtual reality displays, augmented reality displays, and the like. In at least one embodiment, a virtualized environment may be used to generate a virtual interactive display or environment (e.g., a virtual environment) for system users (e.g., doctors, nurses, radiologists, etc.) to interact with. In at least one embodiment, one or more visualization services 1420 may include an internal visualizer, movie, and / or other rendering or image processing capabilities or functions (e.g., ray tracing, rasterization, internal optics, etc.).

[0210] In at least one embodiment, hardware 1322 may include GPU / graphics 1422, AI system 1424, cloud 1426, and / or any other hardware for executing training system 1304 and / or deployment system 1306. In at least one embodiment, GPU / graphics 1422 (e.g., NVIDIA's TESLA and / or QUADRO GPUs) may include any number of GPUs that can be used to perform processing tasks for any feature or functionality of one or more computing services 1416, one or more AI services 1418, one or more visualization services 1420, other services, and / or software 1318. For example, for one or more AI services 1418, GPU / graphics 1422 may be used to perform pre-processing on imaging data (or other data types used by machine learning models), post-processing on the output of machine learning models, and / or perform inference (e.g., to execute machine learning models). In at least one embodiment, cloud 1426, AI system 1424, and / or other components of system 1400 may use GPU / graphics 1422. In at least one embodiment, cloud 1426 may include a GPU-optimized platform for deep learning tasks. In at least one embodiment, AI system 1424 may use a GPU, and cloud 1426 (or at least part of a task being deep learning or reasoning) may be executed using one or more AI systems 1424. Likewise, while hardware 1322 is shown as discrete components, this is not intended to be limiting, and any component of hardware 1322 may be combined with or utilized by any other component of hardware 1322.

[0211] In at least one embodiment, AI system 1424 may include a purpose-built computing system (e.g., a supercomputer or HPC) configured for reasoning, deep learning, machine learning, and / or other artificial intelligence tasks. In at least one embodiment, in addition to CPU, RAM, storage, and / or other components, features, or functions, AI system 1424 (e.g., NVIDIA's DGX) may also include software (e.g., a software stack) that can use multiple GPUs / graphics 1422 to perform GPU-optimized tasks. In at least one embodiment, one or more AI systems 1424 may be implemented in a cloud 1426 (e.g., in a data center) to perform some or all of the AI-based processing tasks of system 1400.

[0212] In at least one embodiment, cloud 1426 may include a GPU-accelerated infrastructure (e.g., NVIDIA's NGC) that may provide a GPU-optimized platform for executing processing tasks of system 1400. In at least one embodiment, cloud 1426 may include an AI system 1424 for executing one or more AI-based tasks of system 1400 (e.g., as a hardware abstraction and scaling platform). In at least one embodiment, cloud 1426 may be integrated with an application coordination system 1428 that utilizes multiple GPUs to enable seamless scaling and load balancing between and among applications and services 1320. In at least one embodiment, cloud 1426 may be responsible for executing at least some services 1320 of system 1400, including one or more compute services 1416, one or more AI services 1418, and / or one or more visualization services 1420, as described herein. In at least one embodiment, cloud 1426 can perform large and small batch inference (e.g., executing NVIDIA's TENSORRT), provide accelerated parallel computing APIs and platforms 1430 (e.g., NVIDIA's CUDA), execute application coordination system 1428 (e.g., Kubernetes), provide graphics rendering APIs and platforms (e.g., for ray tracing, 2D graphics, 3D graphics, and / or other rendering techniques to produce higher quality movie effects), and / or can provide other functionality for system 1400.

[0213] Fig.15A A data flow diagram of a process 1500 for training, retraining, or updating a machine learning model according to at least one embodiment is shown. In at least one embodiment, a non-limiting example may be used. Fig.14The system 1400 of FIG. 1400 performs the process 1500. In at least one embodiment, the process 1500 can utilize services and / or hardware, as described herein. In at least one embodiment, the refined model 1512 generated by the process 1500 can be executed by the deployment system for one or more containerized applications in the deployment pipeline.

[0214] In at least one embodiment, model training 1314 may include retraining or updating an initial model 1504 (e.g., a pre-trained model) using new training data (e.g., new input data (such as customer data set 1506), and / or new ground truth data associated with the input data). In at least one embodiment, to retrain or update the initial model 1504, the output or loss layer of the initial model 1504 may be reset or deleted, and / or replaced with an updated or new output or loss layer. In at least one embodiment, the initial model 1504 may have previously fine-tuned parameters (e.g., weights and / or biases) retained from previous training, so training or retraining 1314 may not take as long or require as much processing as training the model from scratch. In at least one embodiment, during model training 1314, by resetting or replacing the output or loss layer of the initial model 1504, when generating predictions on the new customer data set 1506, the parameters for the new data set may be updated and re-adjusted based on the loss calculation associated with the accuracy of the output or loss layer.

[0215] In at least one embodiment, the pre-trained model 1406 may be stored in a data store or registry. In at least one embodiment, the pre-trained model 1406 may have been trained at least in part at one or more facilities other than the facility performing process 1500. In at least one embodiment, in order to protect the privacy and rights of patients, subjects, or customers of different facilities, the pre-trained model 1506 may have been trained locally using locally generated customer or patient data. In at least one embodiment, the cloud and / or other hardware may be used to train the pre-trained model 1406, but confidential, privacy-protected patient data may not be transmitted to, used by, or accessed by any component of the cloud (or other non-local hardware). In at least one embodiment, if the pre-trained model 1406 is trained using patient data from more than one facility, the pre-trained model 1406 may have been trained separately for each facility before training on patient or customer data from another facility. In at least one embodiment, for example where customer or patient data has been issued privacy issues (e.g., by being relinquished, used for experimental purposes, etc.), or where the customer or patient data is included in a public dataset, customer or patient data from any number of facilities can be used to train pre-trained models 1406 locally and / or externally, such as in a data center or other cloud computing infrastructure.

[0216] In at least one embodiment, when selecting an application to use in a deployment pipeline, a user may also select a machine learning model to use for a particular application. In at least one embodiment, a user may not have a model to use, so the user may select a pre-trained model to use with the application. In at least one embodiment, the pre-trained model may not be optimized to generate accurate results on the customer data set 1506 of the user facility (e.g., based on patient diversity, demographics, type of medical imaging equipment used, etc.). In at least one embodiment, the pre-trained model may be updated, retrained, and / or fine-tuned for use at various facilities before deploying the pre-trained model into a deployment pipeline for use with one or more applications.

[0217] In at least one embodiment, a user may select a pre-trained model to update, retrain, and / or fine-tune, and the pre-trained model may be referred to as the initial model 1504 for the training system in process 1500. In at least one embodiment, a customer data set 1506 (e.g., imaging data, genomic data, sequencing data, or other data types generated by equipment at a facility) may be used to perform model training (which may include, but is not limited to, transfer learning) on ​​the initial model 1504 to generate a refined model 1512. In at least one embodiment, ground truth data corresponding to the customer data set 1506 may be generated by the training system 1304. In at least one embodiment, the ground truth data may be generated, at least in part, at the facility by a clinician, scientist, physician, practitioner.

[0218] In at least one embodiment, AI-assisted annotation can be used in some examples to generate ground truth data. In at least one embodiment, AI-assisted annotation (e.g., implemented using an AI-assisted annotation SDK) can utilize a machine learning model (e.g., a neural network) to generate ground truth data for recommendations or predictions for a customer data set. In at least one embodiment, a user can use the annotation tool within a user interface (graphical user interface (GUI)) on a computing device.

[0219] In at least one embodiment, user 1510 can interact with the GUI via computing device 1508 to edit or fine-tune annotations or automatic annotations. In at least one embodiment, a polygon editing feature can be used to move vertices of a polygon to a more precise or fine-tuned position.

[0220] In at least one embodiment, once the customer data set 1506 has associated ground truth data, the ground truth data (e.g., from AI-assisted annotations, manual labeling, etc.) can be used during model training to generate a refined model 1512. In at least one embodiment, the customer data set 1506 can be applied to the initial model 1504 any number of times, and the ground truth data can be used to update the parameters of the initial model 1504 until an acceptable level of accuracy is achieved for the refined model 1512. In at least one embodiment, once the refined model 1512 is generated, the refined model 1512 can be deployed within one or more deployment pipelines at a facility for use in performing one or more processing tasks with respect to medical imaging data.

[0221] In at least one embodiment, the refined model 1512 can be uploaded to the pre-trained models in the model registry for selection by another facility. In at least one embodiment, this process can be completed at any number of facilities, so that the refined model 1512 can be further refined any number of times on new data sets to generate a more general model.

[0222] Fig. 15B is an example illustration of a client-server architecture 1532 for augmenting an annotation tool with a pre-trained annotation model according to at least one embodiment. In at least one embodiment, an AI-assisted annotation tool 1536 can be instantiated based on the client-server architecture 1532. In at least one embodiment, the AI-assisted annotation tool 1536 in an imaging application can assist radiology physicians, such as identifying organs and anomalies. In at least one embodiment, the imaging application can include a software tool that, as a non-limiting example, helps a user 1510 identify several extreme points on a specific organ of interest in an original image 1534 (e.g., in a 3D MRI or CT scan) and receive automatic annotation results for all 2D slices of the specific organ. In at least one embodiment, the results can be stored in a data store as training data 1538 and used as, for example, but not limited to, ground truth data for training. In at least one embodiment, when the computing device 1508 sends extreme points for AI-assisted annotation, for example, a deep learning model can receive the data as input and return an inference result for segmenting an organ or anomaly. In at least one embodiment, the pre-instantiated annotation tool (e.g., Fig. 15B The AI-assisted annotation tools 1536 in the annotation assistant server 1540 can be enhanced by making an API call (e.g., API call 1544) to a server (such as an annotation assistant server 1540), which may include a set of pre-trained models 1542 stored in, for example, an annotation model registry. In at least one embodiment, the annotation model registry can store pre-trained models 1542 (e.g., machine learning models, such as deep learning models) that are pre-trained to perform AI-assisted annotation on specific organs or abnormalities. In at least one embodiment, these models can be further updated by using a training pipeline. In at least one embodiment, the pre-installed annotation tools can be improved over time as newly labeled data 1312 is added.

[0223] Various embodiments may be described by the following terms:

[0224] 1. A computer-implemented method comprising:

[0225] receiving audio data corresponding to utterance of an utterance;

[0226] computing a weighted vector using a decoder and a transformer-based audio encoder, the weighted vector indicating a plurality of features associated with the audio data;

[0227] using the weighted vector and one or more component vectors to calculate an animation vector corresponding to one or more locations of one or more feature points associated with the digital character representation; and

[0228] The digital character representation is rendered based at least on the animation vectors.

[0229] 2. The computer-implemented method of clause 1, wherein the transformer-based audio encoder is a pre-trained audio encoder, the method further comprising:

[0230] The decoder is trained based at least on the transformer based audio encoder, wherein parameters of the transformer based audio encoder are locked while training the decoder.

[0231] 3. A computer-implemented method as described in clause 1, wherein the plurality of features are selected during training.

[0232] 4. The computer-implemented method of clause 1, further comprising:

[0233] receiving respective layer vectors of layers of the transformer-based audio encoder associated with the plurality of features;

[0234] Determine layer weights for each layer;

[0235] applying the layer weights to the corresponding individual layers; and

[0236] The weighted vector is determined.

[0237] 5. The computer-implemented method of clause 1, wherein the audio data has a duration that is less than a threshold duration.

[0238] 6. The computer-implemented method of clause 5, wherein the plurality of features are determined using a converter layer.

[0239] 7. A computer-implemented method as described in clause 1, wherein the one or more feature points correspond to at least one of a facial feature, a tongue position, an eye position, or a limb position.

[0240] 8. A computer-implemented method as described in clause 1, wherein the plurality of features are extracted from the audio data via processing using a convolutional neural network (CNN).

[0241] 9. The computer-implemented method of clause 1, wherein the component vector comprises at least one of a sentiment vector or a style vector.

[0242] 10. The computer-implemented method of clause 1, wherein the decoder ignores information from one or more previous frames.

[0243] 11. The computer-implemented method of clause 10, further comprising:

[0244] When the volume of the audio data is less than a volume threshold, motion between adjacent frames is penalized.

[0245] 12. A processor, comprising:

[0246] One or more processing units, the one or more processing units being configured to:

[0247] computing a weighted feature vector associated with the audio data using a transformer-based audio encoder and based at least on the audio data corresponding to the utterance;

[0248] calculating position data of one or more feature points of one or more deformable body parts of the virtual character using the weighted feature vector and a component vector indicating one or more characteristics associated with the utterance; and

[0249] For one or more time points in the sequence of time points of the audio data, image data representing the virtual character is rendered based at least on the position data to generate an animation of the character appearing to speak the speech.

[0250] 13. A processor as described in clause 12, wherein the weighted feature vector is based at least on corresponding layer vectors of respective layers of the transformer-based audio encoder, wherein respective layer vectors are associated with a plurality of features extracted from the audio data.

[0251] 14. The processor of clause 12, wherein parameters of the transformer-based audio encoder are locked during a training process of an associated decoder.

[0252] 15. The processor of clause 12, wherein the component vector comprises at least one of a sentiment vector or a style vector.

[0253] 16. The processor of clause 12, wherein the processor is included in at least one of:

[0254] A system for performing simulation operations;

[0255] Systems for performing simulated operations to test or validate autonomous machine applications;

[0256] Systems for performing digital twin operations;

[0257] A system for performing light transport simulations;

[0258] Systems for rendering graphical output;

[0259] Systems for performing deep learning operations;

[0260] Systems implemented using edge devices;

[0261] Systems for generating or presenting virtual reality (VR) content;

[0262] Systems for generating or presenting augmented reality (AR) content;

[0263] Systems for generating or presenting mixed reality (MR) content;

[0264] A system comprising one or more virtual machines VM;

[0265] A system for performing operations of a conversational AI application;

[0266] A system for performing operations of a generative AI application;

[0267] A system for performing operations using a language model;

[0268] A system for performing one or more generative content operations using a large language model (LLM);

[0269] A system implemented at least in part in a data center;

[0270] A system for performing hardware testing using simulation;

[0271] A system for performing one or more generative content operations using a language model;

[0272] Systems for synthetic data generation;

[0273] A collaborative content creation platform for 3D assets; or

[0274] A system implemented at least in part using cloud computing resources.

[0275] 17. A system comprising:

[0276] One or more processing units, the one or more processing units being used to generate animation of the character using position data representing one or more positions of one or more feature points of the character, the position data being calculated at least in part based on a transformer-based audio encoder, the transformer-based audio encoder processing audio data representing speech and component data indicating one or more values ​​corresponding to at least one of a style parameter or an emotion parameter associated with the speech.

[0277] 18. The system of clause 17, wherein the transformer-based audio encoder computes weighted feature vectors based at least on respective layer vectors of respective layers of the transformer-based audio encoder.

[0278] 19. The system of clause 17, wherein parameters of the transformer-based audio encoder are locked during a training process of an associated decoder.

[0279] 20. The system of clause 17, wherein the system is comprised of at least one of:

[0280] A system for performing simulation operations;

[0281] Systems for performing simulated operations to test or validate autonomous machine applications;

[0282] Systems for performing digital twin operations;

[0283] A system for performing light transport simulations;

[0284] Systems for rendering graphical output;

[0285] Systems for performing deep learning operations;

[0286] Systems implemented using edge devices;

[0287] Systems for generating or presenting virtual reality (VR) content;

[0288] Systems for generating or presenting augmented reality (AR) content;

[0289] Systems for generating or presenting mixed reality (MR) content;

[0290] A system comprising one or more virtual machines VM;

[0291] A system for performing operations of a conversational AI application;

[0292] A system for performing operations of a generative AI application;

[0293] A system for performing operations using a language model;

[0294] A system for performing one or more generative content operations using a large language model (LLM);

[0295] A system implemented at least in part in a data center;

[0296] A system for performing hardware testing using simulation;

[0297] A system for performing one or more generative content operations using a language model;

[0298] Systems for synthetic data generation;

[0299] A collaborative content creation platform for 3D assets; or

[0300] A system implemented at least in part using cloud computing resources.

[0301] Other variations are within the spirit of the present disclosure. Thus, while the disclosed technology is susceptible to various modifications and alternative constructions, certain illustrated embodiments thereof are shown in the drawings and have been described in detail above. However, it should be understood that there is no intention to limit the disclosure to one or more specific forms disclosed, but on the contrary, it is intended to cover all modifications, alternative constructions, and equivalents that fall within the spirit and scope of the present disclosure as defined by the appended claims.

[0302] Unless otherwise noted or clearly contradictory to the context, in the context of describing the disclosed embodiments (particularly in the context of the appended claims), the use of the terms "one" and "an" and "the" and similar references should be interpreted as covering the singular and plural, rather than as definitions of terms. Unless otherwise noted, the terms "include", "have", "include" and "contain" should be interpreted as open terms (meaning "including but not limited to") unless otherwise noted. The term "connected" (which refers to a physical connection when unmodified) should be interpreted as partially or completely included, attached to or connected together, even if there are some interventions. Unless otherwise noted herein, references to numerical ranges herein are intended only to be used as a shorthand method of referring to each individual value falling within the range, respectively, and each individual value is incorporated into the specification as if it were individually described herein. Unless otherwise noted or contradictory to the context, the use of the term "set" (e.g., "item set") or "subset" should be interpreted as a non-empty set including one or more members. Furthermore, unless otherwise indicated or contradicted by context, the term "subset" of a corresponding set does not necessarily mean a proper subset of the corresponding set, but rather a subset and a corresponding set may be equivalent.

[0303] Unless expressly indicated otherwise or clearly contradicted by context, conjunctions such as phrases of the form "at least one of A, B, and C" or "at least one of A, B and C" are understood in context to be generally used to refer to an item, clause, or the like that may be A or B or C, or any non-empty subset of the set A and B and C. For example, in the illustrative example of a set having three members, the conjunction phrases "at least one of A, B, and C" and "at least one of A, B and C" refer to any of the following sets: {A}, {B}, {C}, {A, B}, {A, C}, {B, C}, {A, B, C}. Thus, such conjunction language is not generally intended to imply that certain embodiments require the presence of at least one of A, at least one of B, and at least one of C. In addition, unless expressly indicated otherwise or contradicted by context, the term "plurality" refers to a plural state (e.g., "plurality of items" means a plurality of items). The number of items in a plurality of items is at least two, but may be more if expressly indicated or indicated by context. Further, the phrase "based on" means "based at least in part on" rather than "based solely on" unless otherwise specified or clear from context.

[0304] Unless otherwise indicated herein or clearly contradictory to the context, the operations of the processes described herein may be performed in any suitable order. In at least one embodiment, processes such as those described herein (or variations and / or combinations thereof) are performed under the control of one or more computer systems configured with executable instructions and are implemented as code (e.g., executable instructions, one or more computer programs, or one or more applications) that are jointly executed on one or more processors by hardware or a combination thereof. In at least one embodiment, the code is stored on a computer-readable storage medium in the form of, for example, a computer program that includes a plurality of instructions that can be executed by one or more processors. In at least one embodiment, the computer-readable storage medium is a non-transitory computer-readable storage medium that excludes transient signals (e.g., propagated transient electrical or electromagnetic transmissions), but includes non-transitory data storage circuits (e.g., buffers, caches, and queues). In at least one embodiment, code (e.g., executable code or source code) is stored on a set of one or more non-transitory computer-readable storage media (or other memory for storing executable instructions) on which executable instructions are stored, which, when executed by one or more processors of a computer system (i.e., as a result of being executed), causes the computer system to perform the operations described herein. In at least one embodiment, a set of non-transitory computer-readable storage media includes a plurality of non-transitory computer-readable storage media, and one or more of the individual non-transitory storage media in the plurality of non-transitory computer-readable storage media lacks all the code, but the plurality of non-transitory computer-readable storage media stores all the code together. In at least one embodiment, the executable instructions are executed so that different instructions are executed by different processors, for example, a non-transitory computer-readable storage medium stores instructions, and a main central processing unit ("CPU") executes some instructions, while a graphics processing unit ("GPU") executes other instructions. In at least one embodiment, different components of the computer system have separate processors, and different processors execute different subsets of instructions.

[0305] Thus, in at least one embodiment, a computer system is configured to implement one or more services that individually or collectively perform the operations of the processes described herein, and such a computer system is configured with applicable hardware and / or software that enables the implementation of the operations. In addition, a computer system that implements at least one embodiment of the present disclosure is a single device, and in another embodiment is a distributed computer system that includes multiple devices that operate in different ways, so that the distributed computer system performs the operations described herein, and so that a single device does not perform all operations.

[0306] The use of any and all examples or exemplary language (e.g., "such as") provided herein is intended only to better illustrate embodiments of the present disclosure and does not limit the scope of the disclosure unless otherwise required. No language in the specification should be construed as indicating that any non-claimed element is essential to practicing the disclosure.

[0307] All references, including publications, patent applications, and patents, cited herein are hereby incorporated by reference to the same extent as if each reference were individually and specifically indicated to be incorporated by reference and were set forth in its entirety herein.

[0308] In the specification and claims, the terms "coupled" and "connected," as well as their derivatives, may be used. It should be understood that these terms may not be intended as synonyms for each other. On the contrary, in specific examples, "connected" or "coupled" may be used to indicate that two or more elements are in direct or indirect physical or electrical contact with each other. "Coupled" may also mean that two or more elements are not in direct contact with each other, but still cooperate or interact with each other.

[0309] Unless explicitly stated otherwise, it is to be understood that throughout the specification, terms such as “processing”, “computing”, “calculating”, “determining” and the like refer to the actions and / or processes of a computer or computing system or similar electronic computing device that processes and / or converts data represented as physical quantities (e.g., electronic) in registers and / or memories of the computing system into other data similarly represented as physical quantities in the memories, registers or other such information storage, transmission or display devices of the computing system.

[0310] In a similar manner, the term "processor" may refer to any device or part of a memory that processes electronic data from registers and / or memory and converts the electronic data into other electronic data that can be stored in registers and / or memory. As a non-limiting example, a "processor" may be a CPU or a GPU. A "computing platform" may include one or more processors. As used herein, a "software" process may include, for example, software and / or hardware entities that perform work over time, such as tasks, threads, and intelligent agents. Similarly, each process may refer to multiple processes to execute instructions sequentially or in parallel, continuously or intermittently. The terms "system" and "method" may be used interchangeably herein, as long as a system may embody one or more methods, and a method may be considered a system.

[0311] In this document, reference may be made to obtaining, acquiring, receiving or inputting analog or digital data into a subsystem, a computer system or a computer-implemented machine. It is possible to obtain, acquire, receive or input analog and digital data in a variety of ways, such as by receiving data as a parameter of a function call or a call to an application programming interface. In some implementations, the process of obtaining, acquiring, receiving or inputting analog or digital data can be accomplished by transmitting data via a serial or parallel interface. In another implementation, the process of obtaining, acquiring, receiving or inputting analog or digital data can be accomplished by transmitting data from a providing entity to an acquiring entity via a computer network. It is also possible to reference to providing, outputting, transmitting, sending or presenting analog or digital data. In various examples, the process of providing, outputting, transmitting, sending or presenting analog or digital data can be accomplished by transmitting data as an input or output parameter of a function call, an application programming interface or an interprocess communication mechanism.

[0312] Although the above discussion sets forth example implementations of the described techniques, other architectures may be used to implement the described functionality and are intended to fall within the scope of the present disclosure. In addition, although specific responsibilities are defined above for discussion purposes, various functions and responsibilities may be allocated and divided in different ways, depending on the circumstances.

[0313] Furthermore, although the subject matter has been described in language specific to structural features and / or methodological acts, it should be understood that the subject matter claimed in the appended claims is not necessarily limited to the specific features or acts described. Rather, the specific features and acts are disclosed as exemplary forms of implementing the claims.

Claims

1. A computer-implemented method comprising: receiving audio data corresponding to utterance of an utterance; computing a weighted vector using a decoder and a transformer-based audio encoder, the weighted vector indicating a plurality of features associated with the audio data; using the weighted vector and one or more component vectors to calculate an animation vector corresponding to one or more locations of one or more feature points associated with the digital character representation; as well as The digital character representation is rendered based at least on the animation vectors.

2. The computer-implemented method of claim 1 , wherein the transformer-based audio encoder is a pre-trained audio encoder, the method further comprising: The decoder is trained based at least on the transformer based audio encoder, wherein parameters of the transformer based audio encoder are locked while training the decoder.

3. The computer-implemented method of claim 1, wherein the plurality of features are selected during training.

4. The computer-implemented method of claim 1 , further comprising: receiving respective layer vectors of layers of the transformer-based audio encoder associated with the plurality of features; Determine layer weights for each layer; Applying the layer weights to the corresponding layers; as well as The weighted vector is determined.

5. The computer-implemented method of claim 1, wherein the audio data has a duration that is less than a threshold duration.

6. The computer-implemented method of claim 5, wherein the plurality of features are determined using a transformer layer.

7. The computer-implemented method of claim 1, wherein the one or more feature points correspond to at least one of a facial feature, a tongue position, an eye position, or a limb position.

8. The computer-implemented method of claim 1, wherein the plurality of features are extracted from the audio data via processing using a convolutional neural network (CNN).

9. The computer-implemented method of claim 1, wherein the component vector comprises at least one of a sentiment vector or a style vector.

10. The computer-implemented method of claim 1, wherein the decoder ignores information from one or more previous frames.

11. The computer-implemented method of claim 10, further comprising: When the volume of the audio data is less than a volume threshold, motion between adjacent frames is penalized.

12. A processor, comprising: One or more processing units, the one or more processing units being configured to: computing, using a transformer-based audio encoder and based on at least the audio data corresponding to the utterance, a weighted feature vector associated with the audio data; calculating position data of one or more feature points of one or more deformable body parts of the virtual character using the weighted feature vector and a component vector indicating one or more characteristics associated with the utterance; as well as For one or more time points in the sequence of time points of the audio data, image data representing the virtual character is rendered based at least on the position data to generate an animation of the character appearing to speak the speech.

13. The processor of claim 12, wherein the weighted feature vector is based at least on corresponding layer vectors of respective layers of the transformer-based audio encoder, wherein respective layer vectors are associated with a plurality of features extracted from the audio data.

14. The processor of claim 12, wherein parameters of the transformer-based audio encoder are locked during a training process of an associated decoder.

15. The processor of claim 12, wherein the component vector comprises at least one of a sentiment vector or a style vector.

16. The processor of claim 12, wherein the processor is included in at least one of: A system for performing simulation operations; Systems for performing simulated operations to test or validate autonomous machine applications; Systems for performing digital twin operations; A system for performing light transport simulations; Systems for rendering graphical output; Systems for performing deep learning operations; Systems implemented using edge devices; Systems for generating or presenting virtual reality (VR) content; Systems for generating or presenting augmented reality (AR) content; Systems for generating or presenting mixed reality (MR) content; A system comprising one or more virtual machines VM; A system for performing operations of a conversational AI application; A system for performing operations of a generative AI application; A system for performing operations using a language model; A system for performing one or more generative content operations using a large language model (LLM); A system implemented at least in part in a data center; A system for performing hardware testing using simulation; A system for performing one or more generative content operations using a language model; Systems for synthetic data generation; A collaborative content creation platform for 3D assets; or A system implemented at least in part using cloud computing resources.

17. A system comprising: One or more processing units, the one or more processing units being used to generate animation of the character using position data representing one or more positions of one or more feature points of the character, the position data being calculated at least in part based on a converter-based audio encoder, the converter-based audio encoder processing audio data representing speech and component data indicating one or more values ​​corresponding to at least one of a style parameter or an emotion parameter associated with the speech.

18. The system of claim 17, wherein the transformer-based audio encoder computes weighted feature vectors based at least on corresponding layer vectors of respective layers of the transformer-based audio encoder.

19. The system of claim 17, wherein parameters of the transformer-based audio encoder are locked during a training process of an associated decoder.

20. The system of claim 17, wherein the system is included in at least one of: A system for performing simulation operations; Systems for performing simulated operations to test or validate autonomous machine applications; Systems for performing digital twin operations; A system for performing light transport simulations; Systems for rendering graphical output; Systems for performing deep learning operations; Systems implemented using edge devices; Systems for generating or presenting virtual reality (VR) content; Systems for generating or presenting augmented reality (AR) content; Systems for generating or presenting mixed reality (MR) content; A system comprising one or more virtual machines VM; A system for performing operations of a conversational AI application; A system for performing operations of a generative AI application; A system for performing operations using a language model; A system for performing one or more generative content operations using a large language model (LLM); a system implemented at least in part in a data center; A system for performing hardware testing using simulation; A system for performing one or more generative content operations using a language model; Systems for synthetic data generation; A collaborative content creation platform for 3D assets; or A system implemented at least in part using cloud computing resources.

Images