Unsupervised domain adaptation of neural networks

Abstract

Unsupervised domain adaptation of neural networks is disclosed. The approach presented herein provides unsupervised domain transfer learning. In particular, three neural networks can be trained using at least labeled data from a first domain and unlabeled data from a second domain together. A feature extraction network is used to extract features of the data. A first classifier network uses these features to classify the data, while a second classifier network uses these features to determine the relevant domain. A combined loss function is used to optimize the networks with the goal that the feature extraction network extracts features that the first classifier network can use to accurately classify the data, but which prevent the second classifier from determining the domain of the image. This optimization enables object classification to be performed with high accuracy in any domain, even if the second domain can have little labeled training data.

Description

[0001] Cross Reference to Related Applications

[0002] This application claims priority to U.S. Provisional Patent Application Serial No. 63 / 086,544, filed October 1, 2020, entitled “Unsupervised Domain Adaptation of Neural Networks,” the entire contents of which are incorporated herein and used for all purposes. BACKGROUND

[0003] The ability to learn new concepts and skills from generic data and transfer these concepts and skills to similar scenarios is crucial in many modern applications, e.g., that can involve machine learning. For example, it is often the case that a learner has access only to a small subset of unlabeled data on a given domain of interest (e.g., an environment or a type of scenario). However, the learner has access to a larger labeled dataset that can be used for the same task but in a different domain than the target domain. If the difference between the two domains (often referred to as “domain gap”) is not very large, then a model can be trained using the labeled data and the unlabeled data and generalized to the target dataset. This scenario can involve unsupervised domain adaptation. However, traditional unsupervised domain adaptation methods have a problem in that the distribution of the training examples is different from the distribution of the examples used at test or inference in various practical applications. Traditional domain adaptation methods determine the generalization bound based on divergence measures that are difficult to optimize in practice. This leads to a large disconnect between theory and state-of-the-art methods. BRIEF DESCRIPTION OF DRAWINGS

[0004] Various embodiments according to the present disclosure will be described with reference to the drawings, in which:

[0005] Figure 1 An image showing objects in different domains is shown in accordance with at least one embodiment;

[0006] Figure 2 System components for performing unsupervised domain adaptation based training of a network are shown in accordance with at least one embodiment;

[0007] Figure 3 Components of a data classifier system are shown in accordance with at least one embodiment;

[0008] Figure 4 A data graph from multiple domains is shown in accordance with at least one embodiment;

[0009] Figure 5A 、 Figure 5B and Figure 5C Processes for training neural networks and using these neural networks for inference are shown in accordance with at least one embodiment;

[0010] Figure 6System components for performing training and / or inference according to at least one embodiment are shown;

[0011] Figure 7A The inference and / or training logic according to at least one embodiment is illustrated;

[0012] Figure 7B The inference and / or training logic according to at least one embodiment is illustrated;

[0013] Figure 8 An example data center system according to at least one embodiment is shown;

[0014] Figure 9 A computer system according to at least one embodiment is shown;

[0015] Figure 10 A computer system according to at least one embodiment is shown;

[0016] Figure 11 At least a portion of a graphics processor according to one or more embodiments is shown;

[0017] Figure 12 At least a portion of a graphics processor according to one or more embodiments is shown;

[0018] Figure 13 This is an example data flow diagram of an advanced computing pipeline according to at least one embodiment;

[0019] Figure 14 This is a system diagram of an example system for training, tuning, instantiating, and deploying machine learning models in an advanced computing pipeline, according to at least one embodiment; and

[0020] Figure 15A and Figure 15B A data flow diagram of a process for training a machine learning model according to at least one embodiment is shown, as well as a client-server architecture for enhancing annotation tools using a pre-trained annotation model. Detailed Implementation

[0021] Methods according to various embodiments can provide for training or transfer learning across multiple domains. In particular, a neural network can be trained using unsupervised domain adaptation, making it effectively domain-agnostic. An example training method uses three neural networks, including an encoder network and two classifier or discriminator networks. Using object recognition as an example, an encoder network can be trained using labeled images from a first domain and unlabeled images from a second domain to extract features related to object types and encode those features into a latent space. A first object classifier network can utilize those features to attempt to classify the type of object. A second domain classifier network can use those features to attempt to classify or infer the domain in which the image was captured, or from which the features were extracted. The three networks can be trained together and optimized using a common loss function to attempt to arrive at a result in which the object classifier is accurate but the domain classifier is unable to determine the domain. In this way, the encoder network learns to encode domain-agnostic features so that the object classifier can generate accurate classifications for images from any domain. This approach can be extended to any type of task that reasons over data in two or more domains.

[0022] Figure 1 A pair of images including representations of physical objects in two different domains are shown that can be processed according to various embodiments. The first image 100 can include labels for certain types of objects, such as vehicles 102 and people 104, that are to be identified from real-time data when reasoning. These labels are associated with location information, such as bounding boxes, that indicate the portion of the image in which the pixel data for each object is located. The image 100, along with the labels and location information, can be used as training data to train one or more neural networks. The labels and location information enable the image to serve as a source of ground truth data for training.

[0023] As shown, the image can be from a set of similarly labeled images captured from a similar scene or environment. For example, the images can all be taken in urban areas in North America. As such, there can be similar types of objects in the background or elsewhere in the image that can relate to a similar style of buildings, signs, sidewalks, etc. A network trained on these images can learn to extract relevant features from images captured or generated in such a domain. However, it can be desirable to use the network to analyze images from a different domain that can include different types of objects. Figure 1One such image 150 is shown, which can still include one or more types of objects of interest, such as vehicles 152, that will be recognized by the network. However, as shown, other objects in this domain can be very different from those in the first domain. Here, the domain can correspond to the hills or Alps of Europe, rather than the urban environment of North America, where there can be objects not represented in the first domain, such as snow-covered mountains, mountain huts, farm animals, such as sheep, and different versions of certain types of objects, such as different types of signs, roads, or streetlights. There can be other differences as well, such as vehicles being on both sides of the road, if the second domain corresponds to the English countryside, for example. Various other differences can exist for images captured in different domains, as well as non-image data or content in various domains, within the scope of various embodiments.

[0024] It can be desirable to use the same network or model to process data for both domains, such as for a vehicle navigation or assistance system installed in a vehicle model that can be driven in locations associated with either domain. However, as in the case of domain transfer learning, the accuracy of the network will generally not be as good in the second domain, because the training data in this example is only for the first domain, and the second domain is not (or not sufficiently) labeled with training data or ground truth data.

[0025] Methods in accordance with various embodiments can provide for unsupervised domain transfer-based learning of neural networks, models, or algorithms, with results that are similar across multiple domains, where sufficient labeled training data is only available for a subset of the domains. Figure 2 Example components of one such system are shown. In this example, a database 202 of labeled ground truth data is provided for training, where the labeled training data can correspond to data for at least a first domain. At least a subset of these images are provided as training images 204 for training one or more neural networks. In some embodiments, a portion of the labeled ground truth can be reserved for testing purposes. In this example, selected training images 204 are provided as input to a feature extraction network 206. This network can be any suitable network capable of processing input image data and extracting relevant, representative, or unique image features, such as a convolutional neural network (CNN). In this example, the feature extraction network 206 can encode these features into a latent space 208, while in other embodiments, the features can be encoded as a feature vector or latent code, among other such options.

[0026] In this example, the latent space 208 can be provided as input to two different classifier networks, such as an object classifier 210 and an object domain classifier 212 (or per-class domain classifier). When there are multiple classes of data, a per-class domain classifier can be used, and a specialized domain classifier can be used to determine the appropriate domain for each class or object. Other types of networks can also be used for these or other types of data or tasks. In at least one embodiment, each classifier can be a network that operates in the feature space, such as an MLP (multi-level perceptron) fully connected network. More powerful classifiers and discriminators can also be used, but can not be lightweight as this can outweigh any benefits of a slight performance improvement. The object classifier network 210 can analyze the feature data from the latent space and attempt to classify one or more types of objects of interest, or objects of a type for which the network will be trained and for which labeled training data is available. For example, the object classifier can analyze the features of an input training image and identify three different vehicles and two people represented in the image. These results can be provided to an evaluation module 214, which can compare the results to relevant ground truth data to determine any errors in the determination. This error can be provided to another module 216 (or in some embodiments a different part of the same module) to perform tasks such as backpropagation and adjustment of one or more network parameters of the relevant network. These adjusted parameters can then be used with the relevant network for the remainder of the training process. In at least some embodiments, these network parameters can be iteratively updated or modified throughout the training process until an end criterion is met, such as each relevant network converging, reaching a maximum number of training, new training data ending processing, or other such criteria.

[0027] In this example, the features of the latent space 208 are also provided as input to a second classifier network, here the object domain classifier 212. The object domain classifier 212 can analyze the feature data to attempt to determine the domain corresponding to this image, such as a domain corresponding to the environment in which the image data was captured. This can involve distinguishing between two or more specific domains in various embodiments, or can include at least one unknown domain in at least some embodiments. The domain classifier 212 in this example can attempt to determine whether the image data corresponds to a first domain or a second domain. The output determination can also be provided to the evaluation module 214, and error data can be provided to the second module 216 for updating one or more network parameters as appropriate.

[0028] The goal of this training process can be to train the feature extraction network 206 and / or the object classifier 210 to perform equally well on data in the first domain or the second domain. As shown, unlabeled training data 218 can be provided for the second domain, and at least a portion of the images can be selected as training data 204 for training the networks. As with the labeled data in the first domain, the unlabeled data from the second domain will extract features and encode them into the latent space 208, with the object classifier network 210 and the domain classifier network 212 attempting to identify objects and domains, respectively. Errors in these determinations can be used to update the network parameters of the extraction and classifier networks.

[0029] In this example, all three networks are trained together and evaluated using a single loss function that includes terms for each network or task. This approach helps to ensure that the object classifier 210 is able to accurately identify objects from either domain. To help ensure that the object classifier 210 performs equally well on both domains, the feature extraction network can be optimized along with the domain classifier on the same data. The goal of this training process is to train the feature extraction network 206 to extract features that are effectively domain-agnostic, or not specific to a particular domain. In other embodiments, the network can extract domain-related features, but provide lower weight to these features. This can effectively train the feature extraction network to extract features that are only (or at least primarily) related to the object of interest, and not other objects in the image that can be specific to a particular domain, or to extract only types of features that are not typically associated with a particular domain. If the feature extraction network 206 is trained to extract such domain-agnostic features, then the domain classifier network 212 should not be able to identify the domain of a given training image with certainty. In examples, the network will output a value of 0 for the first domain or 1 for the second domain, with intermediate values indicating the relative probability of either domain, and the domain classifier ideally should output a value of 0.5, meaning that the domain classifier is unable to determine the domain, with both domains being equally probable. Training the feature extraction network 206 in this way with the domain classifier 212 helps to train the feature extraction network 206 to extract features that do not have domain specificity, or even domain-indicativeness. Training these features with the object classifier ensures that the extracted features also enable objects to be accurately classified.

[0030] Once the networks have been sufficiently trained, e.g., where all of the networks have converged or the combined loss meets specified loss criteria, at least the extraction and object classifier network parameters can be applied as Figure 3The network of the inference pipeline 300 is shown. In this example, the input image 304 (or image or video data) can be received at a trained feature extraction network 306, which can extract features and encode these features into a latent space 308. The extracted features can be substantially independent of the domain corresponding to the given input image. The latent space 308 can then be provided as input to a trained object classifier 310, which can infer one or more classifications 312 for objects of interest identified in the input image. As previously stated, the accuracy of the inferred classification should not vary by domain, at least not with respect to the domain in which the network was trained. These classifications (including information such as object type and location in the image) can then be provided to an application, system, or service that will use this information to perform one or more tasks.

[0031] The training process presented in this paper enables the feature extraction network to extract features that minimize the influence of the domain. In at least one embodiment, the objective can be to extract features that correspond only to the object of interest and in a manner completely independent of the domain. Choosing an appropriate training loss function helps reduce the importance of the extraction network to the domain, reducing the weight of any features that do not directly correspond to the object of interest or object category. This approach can be used to provide adaptation across several different domains that may exhibit different differences. As mentioned above, the type of domain may vary depending on the content type, such as video, audio, images, data, or other types of content. A domain can correspond to any differences between datasets that may prevent a network trained on one dataset from correctly analyzing it; this could be related to location, origin, object type, style, appearance, etc. Various differences may exist between data from different domains, which may cause a network trained on data from one domain to fail to accurately infer data from another domain. For example, with images, such as... Figure 1 As shown, this could be related to differences in architecture, style, object appearance, object selection, lighting, and other similar variations. Different domains can also represent the same type of objects with similar aspects, but the actual data may differ from the synthetic data or other similar variations. The methods presented here for various embodiments are general, use lightweight models, and are scalable, capable of handling many (if not all) domain differences, and offer high performance levels. Examples of tasks that can leverage these advantages include, but are not limited to, natural language processing, navigation or robotics simulation and data analysis, as well as many other such examples.

[0032] Methods in accordance with various embodiments use a domain adversarial framework for domain adaptation, which can provide practical learning algorithms for use with neural networks. In at least one embodiment, generalization bounds exploit a measure of difference between distributions based on variational characterizations of f-divergences. Such bounds can support various types of divergences that can arise in practice. A general algorithm for domain adversarial learning for full f-divergences is presented. Some f-divergences, including several that have not been considered in domain adversarial learning before, have been shown to achieve state-of-the-art results in practice. Various insights can be provided about how the choice of a particular divergence affects the transfer performance on real-world datasets. By further identifying the optimization problem as a game, such as a Stackelberg game in at least some embodiments, optimizers from game optimization can be leveraged to achieve additional performance improvements in the training algorithm. It has been shown that such an f-domain adversarial framework can achieve state-of-the-art results on challenging datasets such as Office-31 and Office-Home without the need for additional hyperparameters. As described with respect to the graph 400 of FIG. 4, domain adaptation can be performed by a learner (e.g., a neural network) trained on a large amount of labeled data, where the learner is also expected to perform well in a target domain (indicated by “+” in the graph). For example, the decision boundary can correspond to a two-layer neural network trained using a method such as f-DAL discussed in greater detail elsewhere herein. Figure 4 As described with respect to the graph 400 of FIG. 4, domain adaptation can be performed by a learner (e.g., a neural network) trained on a large amount of labeled data, where the learner is also expected to perform well in a target domain (indicated by “+” in the graph). For example, the decision boundary can correspond to a two-layer neural network trained using a method such as f-DAL discussed in greater detail elsewhere herein.

[0033] In at least one embodiment, a framework is utilized to support a full family of f-divergences, which is practical for modern neural networks. In particular, a measure of divergence can be introduced between distributions and their corresponding learned bounds derived therefrom. The difference in this case can be based on variational characterizations of f-divergences, and can include previous theoretical results and practical results. In at least some embodiments, an invariant representation can be learned using any f-divergence, including several divergences that have not been considered in domain adversarial learning before. From an optimization perspective, a Stackelberg equilibrium can be an optimal solution for the framework under mild conditions in at least one embodiment. Such a solution is able to use a minimax optimizer plug-and-play within the framework.

[0034] Methods in accordance with at least one embodiment can be discussed with respect to an unsupervised domain adaptation scenario. During training, it can be assumed that a learner has access to a source dataset of n s labeled examples and a target dataset of n t unlabeled examples where the source inputs are sampled independently and identically from a distribution P s (source distribution) over an input space X, and the target inputs is sampled independently and identically from distribution P t (ground truth distribution) on X. In the case of binary classification, Y = {0, 1}, and in the multi-class classification scenario, Y = {1,..., k}.

[0035] In at least one embodiment, a labeling function can be represented as f : X → Y, and source and target labeling functions are represented as f s and f t , respectively. The task of unsupervised domain adaptation is to find a hypothesis function h : X → Y that generalizes to the target dataset T (i.e., is as errorless as possible when compared to the true labels . The risk of the hypothesis function h with respect to the labeling function f can be defined using a loss function l : Y x Y → R + as follows:

[0036]

[0037] It can also be assumed that l satisfies the triangle inequality. To make the notation simple, one approach can be defined as:

[0038]

[0039]

[0040] where indices S and T represent the source and target domains, respectively. In addition, can be used to refer to the empirical risk of the source dataset S and the target dataset T.

[0041] One important aspect of domain adaptation, at least in some embodiments, involves the difference between the source distribution and the target distribution. This distinguishes the transfer method and more general transfer learning from traditional supervised learning methods. In at least one embodiment, a generalization boundary is derived that captures the entire f-divergence family. A new difference between the source distribution and the target distribution can be defined, at least partially based on the variational feature of various f-divergence choices. These new differences play a fundamentally important role in our work.

[0042] In at least one embodiment, an f-divergence can be defined as let P s and P t represent the distribution functions with densities P s and P t , respectively. This approach can let P s and P t be absolutely continuous with respect to the base measure dx. Let be a convex lower semi-continuous function satisfying . The f-divergence can be defined as:

[0043]

[0044] As for the variational characterization of f-divergence, one can estimate f-divergence from examples by converting the estimation problem into a variational optimization using the general variational method. In one example, any f-divergence can be written as:

[0045]

[0046] where is the (Fenchel) conjugate function of : R + → R defined as and The equality holds if T is the set of all measurable functions. Many popular divergences used heavily in machine learning and information theory are special cases of f-divergence. For simplicity, one can assume and each density (e.g., p s and p t ) is absolutely continuous.

[0047] Domain adaptation methods are generally based on the idea of bounding the gap between the error functions of the source and target domains according to the difference between the probability distributions of the source and target domains. Measuring the similarity between distributions P s and P t is important for deriving generalization bounds and / or algorithm design. In at least one embodiment, it is assumed that the risk of any binary classifier in the class H can be bounded using the following theorem:

[0048] Theorem 1. If l(x, y) = |h(x) - y| and H is a class of functions, then for any h e H:

[0049]

[0050] Here, is TV and T is the set of measurable functions. TV is an f-divergence such that The above. For any function one can replace D TV (P s || P t ) in the above equations with The above theorem bounds the target error of a classifier in terms of the source error, the divergence between the two domains, and the difference in labeling functions. Unfortunately, D TV (P s || P t ) cannot be estimated from finite examples of arbitrary distributions. It is also a very loose upper bound as it involves the supremum over all measurable functions and does not take into account the hypothesis class.

[0051] In at least one embodiment, a new divergence can be utilized that can be used to address at least two of the above problems, namely (1) the estimation of the divergence for a finite number of examples of arbitrary distribution and (2) the restriction on the divergence for a set containing the hypothesis class H. A second definition of the divergence can be provided as follows:

[0052] Let be a convex lower semi-continuous function satisfying the Fenchel conjugate of , let be a set of measurable functions such that:

[0053]

[0054] The divergence between P s and P t can be defined as:

[0055]

[0056] The divergence can be interpreted as a lower bound estimator for the general f-divergence class. Thus, for any hypothesis class H and the selection of , never exceeds its corresponding f-divergence. Its computation can be bounded in terms of the finite examples. Furthermore, if and l(h(x),h'(x)) = 1(h(x)≠h'(x)), the HΔH divergence can be recovered, which corresponds to TV.

[0057] A third definition of the divergence can be provided as follows:

[0058] With the same assumptions as above, the divergence between two distributions P s and P t is defined as:

[0059]

[0060] Taking the upper bound value of over h∈H, we have Thus This bound is useful in deriving practical algorithms. As a lower bound (Lemma 1), for any two functions h, h' in H:

[0061]

[0062] Lemma 1 is the foundation for deriving the divergence-based generalization bound. Specifically, it uses f-divergence to bound the gap between error functions according to the difference between the source and target distributions. It can also be shown that can be estimated from finite examples. For Lemma 2, assume l: Y x Y → [0, 1], and Let S and T be two empirical distributions, containing n data points, drawn from P s and P t independently and identically distributed. Let denote the Rademacher complexity of a given function class, and then there is at least a 1 - δ probability that:

[0063]

[0064] In Lemma 2, it is shown that the empirical converges to the true difference. It can then be estimated using a set of finite examples from both distributions. This difference is bounded by the complexity of the hypothesis class and the number of examples (n).

[0065] In at least one embodiment, using the proposed divergence and results from the previous section, a generalization bound can be used to estimate the error of a classifier in the target domain. Based on the estimation of from finite examples, the generalization Rademacher complexity bound can also be used for binary classifiers. This bound can be shown to generalize previous existing results.

[0066] For Theorem 2 (generalization bound), assume and for any a, b, c e Y, l(a, b) ≤ l(a, c) + l(c, b). Let Let h * be the ideal joint hypothesis. This will result in:

[0067]

[0068] The three terms in this upper bound have similarities with previously proposed bounds, with the notable difference being the use of the reduced estimator of the variational feature of the general f-divergence. This generalizes and includes popular divergences that are commonly used in practice. Intuitively, the first term in the bound accounts for the source error, the second term corresponds to the difference between the marginal distributions, and the third term measures the ideal joint hypothesis (h *If H is sufficiently expressive and similar to the labeling function, the last term can be reduced to a smaller value. The ideal joint hypothesis incorporates the concept of adaptation: successful adaptation cannot be expected when the optimal hypothesis performs poorly in either domain.

[0069] For Theorem 3 (the generalization boundary with Rademacher complexity), let Let S and T be two empirical distributions (i.e., datasets containing n data points from P). s and P t (Independent and identically distributed sampling). (This indicates...) Then there is at least a probability of 1-δ:

[0070]

[0071] In Theorem 3, based on the Rademacher complexity of class H, the generalization boundary of the binary classifier is calculated. It can be seen that, under the ideal joint assumption... Under the assumption that the generalization error can be reduced by jointly minimizing the risk in the source domain, the difference between the two distributions, and normalizing the model to limit the complexity of the hypothesis class.

[0072] The previously proposed theory can be used to derive a generalized domain adversarial learning framework. An important aspect of domain adversarial training is simultaneously minimizing the source error and aligning two distributions in the representation space Z. Specifically, it is assumed that h can be... The combination (i.e., let) and Another function class), where g: X→Z. This can be interpreted as converting two densities p s and p t Proceeding to the mapping of the representation space Z, where the classifier Operation. Therefore, you can refer to... and These are used as advances for the source domain density and the target domain density, respectively. The above discussion... Figure 2 An example of such a framework is given in the document.

[0073] In at least one embodiment, in order for adaptation in representation space Z to be possible, there must be a Make the ideal joint risk λ * It can be ignored. If p z s =p z t This situation may exist. In other words, p z sand p z t The difference between p * should be small, and the ideal joint risk λ * should be negligible. In at least some embodiments, these are sufficient and necessary conditions. A first assumption can be made that there exists g G and such that the ideal joint risk (λ * ) can be made negligible. It can also be assumed that the class conditional distributions between the source and target are similar. Neural networks are capable of learning rich and powerful representations, and in practical scenarios, g and are neural networks.

[0074] According to Theorem 2 and the first assumption, the target risk R can be optimized by jointly minimizing the error in the source domain and the difference between the two distributions. Let y := f s (x), the optimization objective can be written as:

[0075]

[0076] Here, l is a proxy loss function used to minimize the empirical risk in the source domain. It is not necessarily a binary classification loss, however (i.e., it can be a cross-entropy loss). Under certain assumptions (e.g., Proposition 1) and the use of Lemma 1, the minimization problem above can be upper bounded (i.e., replaced) by the following min-max objective:

[0077]

[0078] where the difference between the last two terms can be called d s,t. . This result can now be formalized in Proposition 1. Assume d s,t. takes the form as shown above, where and for any there exists such that for any z supp(p z t (z)), where is the derivative of . The optimal d s,t is then given by (i.e., d ).

[0079] If the feature extractor g G is allowed to be one that minimizes both the source error and the difference term, then the corresponding equation can be rewritten as:

[0080]

[0081] For GANs, The choice of g can be arbitrary. For multi-class scenarios, let where argmax a is the index of the maximum element of vector a. The binary case can utilize This means that the choice of g is The domain of g is where k classes are used for multi-class scenarios, for binary classification. Intuitively, is a helper per-class domain classifier. For the choice of g, one can choose a monotonically increasing function as much as possible.

[0082] As mentioned before, one important goal of this framework can be to have a minimax optimization problem with a desired (i.e., optimal) solution, which under mild assumptions can be a Stackelberg equilibrium. This perspective enables the incorporation of various state-of-the-art optimizers into the framework, such as optimizers for game optimization. Accordingly, additional gradient algorithms and their aggressive versions can be used in the unsupervised domain adaptation framework.

[0083] In at least one embodiment, the existence of a Stackelberg equilibrium can be proven in f-domain adversarial learning (f-DAL). Let G and be a class of functions defined by a fixed parameter functional, such as a neural network with a fixed architecture, and define ω1 such that it is a vector composed of parameters of the feature extractor g and the source classifier Similarly, let ω2 represent the parameters of the helper classifier, Ω1 and Ω2 represent their respective domains. The corresponding equations above can be rewritten as:

[0084]

[0085] In general, V is non-convex in ω1 and non-concave in ω2, and for a minimax game, a Nash equilibrium can not exist. Stackelberg equilibrium is more general than Nash equilibrium, reflecting the sequential nature of the zero-sum game equations described above. It can be proven that the optimal solution of f-DAL is a Stackelberg equilibrium in at least one embodiment. Under the assumption that V(ω1,.) is (locally) strongly concave in ω2, such equilibrium is a stationary point, and then gradient algorithms can be used to search for such ideal solutions. In the following theorem, the forward-propagating significant form can be used to emphasize the dependence on the feature extractor g instead of p z s ,p z t .

[0086] For Theorem 4 (Stackelberg equilibrium, informal), assume d s,t In the above form, and assume there exists an optimal g * ∈ G that maps both the source and target distributions to the same distribution. Also assume there exists an optimal classifier that can produce ground truth in the neighborhood, and for any g ∈ G and There exists such that The goal of adversarial learning for f can have a Stackelberg equilibrium at When ( e.g., in a binary classification scenario), the Stackelberg equilibrium can show as a Nash equilibrium.

[0087] As discussed, in at least one embodiment, the optimal solution in f-DAL is a Stackelberg equilibrium, which is more general than a Nash equilibrium. To converge to a Nash equilibrium, the simplest method can be GDA, but the last iteration of GDA does not converge in the bilinear case. To speed up and stabilize the convergence, an extra-gradient (EG) method can be used. There is an aggressive extra step that can be more stable than the general form of EG and can be more suitable to converge to a Stackelberg equilibrium. To quantify whether exploiting Theorem 4 would bring practical benefits, a larger extrapolation step can be allowed, which is referred to herein as an aggressive extra-gradient (AExG).

[0088] In at least one embodiment, such an algorithm can be implemented in PyTorch, using a ResNet-50 pre-trained on ImageNet as a feature extractor. The main classifier and the auxiliary classifier Both the object classifier and the domain classifier can be two-layer neural networks with Leaky-Relu activation function. Spectral Normalization (SN) can be used for both the main and auxiliary classifiers. Using Leaky-Relu activation function instead of Relu does not show significant improvement for transfer learning. One reason for this is to avoid gradient problems and instability during some of the divergence training of the first epoch. For simplicity, and to be fair to previous work, the updates can be performed simultaneously. For optimization, the example experiments utilize mini-batch (32) SGD (or GDA) with Nesterov momentum. For experiments using AExG, one version of ExtraGradient can be implemented using momentum (0.9). For aggressive steps, a multiplier [10, 1] can be used with a polynomial decay rate, with power = 0.5 for the first 10K iterations. In all cases, the learning rate for the classifiers is set to 10 times the learning rate of the feature extractor (0.01), with the values adjusted accordingly.

[0089] This approach provides a new perspective on the domain adversarial problem by deriving new theory and learning algorithms that support the full family of f-divergences and are applicable to modern neural networks. The learning objectives of this framework can be implemented as a Stackelberg game, using optimizers such as Optimizer for Game Optimization, to achieve additional performance improvements. Large-scale experiments show that any f-divergence can be used to minimize the difference between the source and target domains in the representation space. The research also shows that some of the divergences that were not previously considered in domain adversarial learning achieve results in practice that are at least as good as other state-of-the-art methods, reducing the need for additional techniques and hyperparameter tuning required by these existing methods.

[0090] Figure 5AAn example process 500 for performing unsupervised domain transfer training can be used in accordance with various embodiments is illustrated. It will be appreciated that for this and other processes presented herein, additional, fewer, or alternative steps can be performed in similar or alternative order, or at least partially in parallel, within the scope of various embodiments, unless otherwise specifically noted. Moreover, while image data is used as an example, it will be appreciated that other types of data can be used for such processes as discussed and suggested elsewhere herein. In this example, a set of training data is obtained 502 or generated, including labeled images for a first domain and unlabeled images for a second domain. In other examples, there can be images from multiple domains. In this example, an image is selected 504 from the training data, where the image can correspond to a labeled image from the first domain or an unlabeled image from the second domain. In some embodiments, some amount of pre-processing can be performed, such as adjusting the size, resolution, color depth, format, noise level, or other aspects of the image. The particular image can be processed using a feature extraction 506 network, and the features extracted from the image can be encoded into a latent space. In at least one embodiment, these features should primarily correspond to one or more objects of a particular class or type of interest represented in the image. A first classifier network can be used to analyze 508 the encoded features to classify at least one object of interest in the selected image. A second classifier network can also be used to infer the domain corresponding to the selected image and the predicted object, analyzing 510 the encoded features in parallel. The performance of these three networks can then be evaluated 512 using a loss function that includes at least terms related to the accuracy of the object and domain classifications. Based at least in part on the loss value determined using this loss function, one or more network parameters (or network weights) of one or more of the feature extraction or classifier networks can be adjusted 514 in an attempt to at least maximize the accuracy of the first classifier network, or object classification, and the uncertainty of the second classifier network, such that the second network cannot identify the domain of the selected image with any reasonable certainty. In at least one embodiment, this corresponds to the result of the second classifier network inferring that the image is equally likely to correspond to any of the set of domains for which the network is being trained. Once any adjustments have been made to the network parameters, a decision can be made 516 as to whether at least one final criterion has been met, such as the network meeting a convergence criterion, reaching a maximum number of training iterations or all training data having been processed, among other such options. If not, the process can continue with the next selected training image. If at least one training end criterion is met, the network parameters for at least one of the feature extraction network or the first classifier network can be provided 518 for use in inference.

[0091] Figure 5BAnother example training process 540 that can be used in accordance with various embodiments is shown. In this example, a set of features representative of a classifiable input are encoded 542 into a latent space, with extraction and encoding performed using a first neural network. The input can then be classified 544 based at least in part on the encoded features using a second neural network. In at least one embodiment, this can involve classifying one or more objects or instances in the classifiable input. A domain corresponding to the classifiable input can also be inferred 546 based at least in part on the encoded features using a third neural network. One or more network parameters can then be adjusted 548 for one or more of the neural networks until the first neural network encoding enables the second neural network to accurately classify the classifiable input, but prevents the third neural network from being able to infer the domain of the classifiable input. Such a result means that data can be accurately classified regardless of which domain is associated with that data. This occurs even if training data from a domain is unlabelled, or at least insufficiently labelled to provide accurate domain-specific training. The network parameters for at least the first or second neural network can then be provided for use in inferring one or more tasks involving data from one or more domains.

[0092] Figure 5C An example process 580 for classifying data using a network trained using processes such as those discussed with respect to Figure 5A or Figure 5B The example process 580 classifies data using a network trained using processes such as those discussed with respect to

[0093] As an example, Figure 6An example network configuration 600 is shown that can be used to provide or generate content or other data. In at least one embodiment, a client device 602 can generate content for a session using components of a content application 604 on the client device 602 and data stored locally on the client device. In at least one embodiment, a content application program 624 (e.g., an image generation or editing or analysis application) executing on a content server 620 (e.g., a cloud server or edge server) can initiate a session associated with at least a client device 602, as can utilize a session manager and user data stored in a user database 634, and can cause content 632 to be determined by a content manager 626, sent to the client device 602 using an appropriate transport manager 622, to be sent by download, streaming, or other such transport channel. The content server 620 can also include one or more training modules 630 for training components, networks, or pipelines. The server 620 can include a classifier component 628 for classifying input data or running other inferences, and providing that content or content generated from those instances. In at least one embodiment, a client device 602 receiving this content can provide this content to a corresponding content application program 604 that provides at least some of this content for rendering by the client device 602, such as image or video content through a display 606 and audio content such as sound and music through at least one audio playback device 608, such as speakers or headphones.

[0094] In at least one embodiment, at least some of this content can have been stored, rendered on the client device 602, or accessible to the client device 602, such that at least this portion of content does not require transmission over a network 640, such as the content can have been previously downloaded or stored locally on a hard drive or optical disc. In at least one embodiment, a transmission mechanism such as a data stream can be used to transmit this content from the content server 620 or content database 634 to the client device 602. In at least one embodiment, at least a portion of this content can be obtained or streamed from another source, such as a third party content service 660, which can also include a content application program 662 for generating or providing content. In at least one embodiment, portions of this functionality can be performed using multiple computing devices or multiple processors within one or more computing devices, such as can include a combination of CPUs and GPUs.

[0095] In at least one embodiment, content application 624 includes a content manager 626 that can determine or analyze content prior to transmitting that content to client device 602. In at least one embodiment, content manager 626 can also include or use other components capable of generating, modifying, or enhancing content to be provided. In at least one embodiment, this can include a rendering engine for rendering image or video content. In at least one embodiment, a classification component 628 can be used to infer classifications that can be used by content application 624 on a server or content application 604 on a client device to generate relevant content or perform intended tasks. In at least one embodiment, a training component 630 can be used to train classifier networks, and any related networks or algorithms, using processes such as those described herein. In at least one embodiment, content manager 626 can transmit the content to client device 602. In at least one embodiment, content application 604 on client device 602 can also include components such as classification component 612 or content generation module, such that any or all of this functionality can additionally or alternatively be performed on client device 602. In at least one embodiment, content application 662 on third party content service system 660 can also include such functionality. In at least one embodiment, a location where at least some of this functionality is performed can be configurable, or can depend on factors such as a type of client device 602, or an availability of network connections with appropriate bandwidth. In at least one embodiment, a system for content generation can include any appropriate combination of hardware and software in one or more locations. In at least one embodiment, generated image or video content at one or more resolutions can also be provided to or made available for other client devices 650, such as for download or streaming from a media source that stores a copy of the image or video content. In at least one embodiment, this can include transmitting game content for a multiplayer game, where different client devices can display the content at different resolutions, including one or more super- resolutions.

[0096] In this example, the client devices can include any appropriate computing devices, such as can include desktop computers, notebook computers, set-top boxes, streaming devices, game consoles, smartphones, tablet computers, VR headsets, AR eyewear, wearable computers, or smart televisions. Each client device can submit requests across at least one wired or wireless network, such as can include the Internet, an Ethernet network, a local area network (LAN), or a cellular network, among other such options. In this example, the requests can be submitted to an address associated with a cloud provider, which can operate or control one or more electronic resources in a cloud provider environment, such as can include a data center or a server farm. In at least one embodiment, the requests can be received or processed by at least one edge server that is 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 causing the client devices to interact with servers that are closer in distance, while also improving security of the resources in the cloud provider environment.

[0097] 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 to provide image or video content for testing or validating autonomous machine applications, or to perform deep learning operations. In at least one embodiment, such a system can be implemented using edge devices, or can incorporate one or more virtual machines (VMs). In at least one embodiment, such a system can be implemented at least partially in a data center or at least partially using cloud computing resources.

[0098] Inference and training logic

[0099] Figure 7A Inference and / or training logic 715 are used to perform inferencing and / or training operations associated with one or more embodiments. In at least one embodiment, as shown in FIG. 7, inference and / or training logic 715 are used to perform inferencing and / or training operations associated with inference and / or training logic 310, 410, and / or 510 of FIGS. 3, 4, and / or 5. Inference and / or training logic 715 are used to perform inferencing and / or training operations in conjunction with other Figure 7A and / or Figure 7B Details regarding inference and / or training logic 715 are provided below in conjunction with FIG. 7.

[0100] In at least one embodiment, inference and / or training logic 715 can include, without limitation, code and / or data storage 701 for storing forward and / or output weights and / or input / output data, and / or other parameters of neurons or layers of a neural network configured in aspects of one or more embodiments that are trained and / or used for inferencing. In at least one embodiment, training logic 715 can include or be coupled to code and / or data storage 701 for storing graph code or other software to control timing and / or order, in which weight and / or other parameter information is loaded to configure logic, including integer and / or floating point units (collectively, arithmetic logic units (ALUs)). In at least one embodiment, code, such as graph code, loads weight or other parameter information into processor ALUs based on an architecture of a neural network to which the code corresponds. In at least one embodiment, code and / or data storage 701 stores weight parameters and / or input / output data of each layer of a neural network trained or used in conjunction with one or more embodiments during forward propagation of input / output data and / or weight parameters during training and / or inferencing using aspects of one or more embodiments. In at least one embodiment, any portion of code and / or data storage 701 can be included with other on-chip or off-chip data storage, including a processor’s Ll, L2, or L3 cache or system memory.

[0101] In at least one embodiment, any portions of code and / or data storage 701 can 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 can be cache memory, dynamic random addressable memory (“DRAM”), static random addressable memory (“SRAM”), non-volatile memory such as flash memory, or other storage. In at least one embodiment, a choice of whether code and / or data storage 701 is internal or external to a processor, e.g., or comprised of DRAM, SRAM, flash, or some other storage type, can depend on available storage space on-chip or off-chip, latency requirements of training and / or inferencing functions being performed, batch size of data used in inferencing and / or training of a neural network, or some combination of these factors. In at least one embodiment, inferencing and / or training logic 715 can include, without limitation, code and / or data storage 705 to store backward and / or output weight and / or input / output data corresponding to neurons or layers of a neural network being trained and / or used for inferencing in aspects of one or more embodiments. In at least one embodiment, code and / or data storage 705 stores weight parameters and / or input / output data for each layer of a neural network being trained or used in conjunction with one or more embodiments during backpropagation of input / output data and / or weight parameters during training and / or inferencing using aspects of one or more embodiments. In at least one embodiment, training logic 715 can include or be coupled to code and / or data storage 705 to store graph code or other software to control timing and / or order, wherein weight and / or other parameter information is loaded to configure logic, including integer and / or floating point units (collectively, arithmetic logic units (ALUs)).

[0102] In at least one embodiment, code such as graph code loads weight or other parameter information into processor ALUs based on an architecture of a neural network to which the code corresponds. In at least one embodiment, any portion of code and / or data storage 705 can be included with other on-chip or off-chip data storage, including a processor’s LI, L2, or L3 cache or system memory. In at least one embodiment, any portion of code and / or data storage 705 can 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 can be cache memory, DRAM, SRAM, non-volatile memory (e.g., Flash) or other storage. In at least one embodiment, a choice of whether code and / or data storage 705 is internal or external to a processor, e.g., is made up of DRAM, SRAM, Flash, or some other storage type, depends on whether available storage is on-chip or off-chip, latency requirements of training and / or inferencing functions being performed, batch size of data being used in inference and / or training of a neural network, or some combination of these factors.

[0103] In at least one embodiment, code and / or data storage 701 and code and / or data storage 705 can be separate storage structures. In at least one embodiment, code and / or data storage 701 and code and / or data storage 705 can be the same storage structure. In at least one embodiment, code and / or data storage 701 and code and / or data storage 705 can 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 can be included with other on-chip or off-chip data storage, including a processor’s LI, L2, or L3 cache or system memory.

[0104] In at least one embodiment, inference and / or training logic 715 can include, without limitation, one or more arithmetic logic unit(s) (“ALUs”) 710 (including integer and / or floating point units) for performing logical and / or mathematical operations, based at least in part on training and / or inference code (e.g., graph code) or instructions by which they are instructed to operate, whose results can be activations (e.g., output values from layers or neurons within a neural network) stored in an activation storage 720 that are functions 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 stored in activation storage 720 are generated by ALU 710 executing linear algebraic and / or matrix-based mathematics in response to executing instructions or other code, where weight values stored in code and / or data storage 705 and / or 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 can be stored in code and / or data storage 705 or code and / or data storage 701 or other on-chip or off-chip storage.

[0105] 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 located outside the processor or other hardware logic device or the circuitry 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 the execution unit of the processor, which may be within the same processor or distributed among 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 circuitry, while in another embodiment, they may be in different processors or other hardware logic devices or circuitries, or in some combination of the same and different processors or other hardware logic devices or circuitries. In at least one embodiment, any portion of activation storage 720 may be included together with other on-chip or off-chip data storage, including the processor's L1, L2, or L3 cache or system memory. Furthermore, the inference and / or training code may be stored alongside other code accessible to the 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. In at least one embodiment, the active memory 720 may be a cache memory, DRAM, SRAM, non-volatile memory (e.g., flash memory), or other storage. In at least one embodiment, the active memory 720 may be wholly or partially located within or outside one or more processors or other logic circuitry. In at least one embodiment, the choice of whether the active memory 720 is internal to or external to the processor may depend on the availability of on-chip or off-chip storage, the latency requirements for training and / or inference functions, the batch size of data used in inference and / or training the neural network, or some combination of these factors, for example, or may include DRAM, SRAM, flash memory, or other storage types.

[0106] In at least one embodiment, Figure 7A The inference and / or training logic 715 shown can be used in conjunction with an application-specific integrated circuit (“ASIC”), such as those from Google. Processing unit, from Graphcore TM Inference processing unit (IPU) or from Intel Corp. (e.g., "LakeCrest") processor. In at least one embodiment, Figure 7AThe inference and / or training logic 715 shown can be used in conjunction with central processing unit (“CPU”) hardware, graphics processing unit (“GPU”) hardware, or other hardware (such as field programmable gate array (“FPGA”)).

[0107] Figure 7B Inference and / or training logic 715 according to at least one or more embodiments is illustrated. In at least one embodiment, the inference and / or training logic 715 may include, but is not limited to, hardware logic, wherein computational 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 can be used in conjunction with an application-specific integrated circuit (ASIC), such as those from Google. Processing unit, from Graphcore TM Inference processing unit (IPU) or from Intel Corp. (e.g., "LakeCrest") processor. In at least one embodiment, Figure 7B The inference and / or training logic 715 shown 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 inference 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, each of code and / or data storage 701 and code and / or data storage 705 is associated with dedicated computing resources (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) only on the information stored in code and / or data storage 701 and code and / or data storage 705, respectively, and the results of the function execution are stored in activation storage 720.

[0108] In at least one embodiment, each of code and / or data stores 701 and 705 and corresponding compute hardware 702 and 706, respectively, correspond to different layers of a neural network, such that activations resulting from one “storage / compute pair 701 / 702” of code and / or data store 701 and compute hardware 702 provide input to the next “storage / compute pair 705 / 706” of code and / or data store 705 and compute hardware 706 in order to reflect the conceptual organization of a 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 inference and / or training logic 715 after or in parallel with storage compute pairs 701 / 702 and 705 / 706.

[0109] Data center

[0110] Figure 8 An example data center 800 that can use at least one embodiment is shown. In at least one embodiment, data center 800 includes a data center infrastructure layer 810, a framework layer 820, a software layer 830 and an application layer 840.

[0111] In at least one embodiment, as shown Figure 8 Data center infrastructure layer 810 can include resource orchestrators 812, grouped computing resources 814 and node computing resources (“node C.R.s”) 816(1)-816(N), where “N” represents any positive integer. In at least one embodiment, node C.R.s 816(1)-816(N) can include, but are not limited to, any number of central processing units (“CPUs” or “processors”) including accelerators, field programmable gate arrays (FPGAs), graphics processors, etc., memory devices (e.g., dynamic random access memory), storage devices (e.g., solid state 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 node C.R.s 816(1)-816(N) can be a server having one or more of above-described computing resources.

[0112] In at least one embodiment, grouped computing resources 814 can include individual groupings of nodes C.R. housed within one or more racks (not shown), or housed within a number of racks (also not shown) within various geographic locations. Individual groupings of nodes C.R. within grouped computing resources 814 can include grouped computing, network, memory, or storage resources that can be configured or allocated to support one or more workloads. In at least one embodiment, several nodes C.R. including CPUs or processors can be grouped within one or more racks to provide computing resources to support one or more workloads. In at least one embodiment, one or more racks can also include any number of power modules, cooling modules, and network switches in any combination.

[0113] In at least one embodiment, resource orchestrator 812 can configure or otherwise control one or more nodes C.R. 816(1)-816(N) and / or grouped computing resources 814. In at least one embodiment, resource orchestrator 812 can include a software design infrastructure (“SDI”) management entity for data center 800. In at least one embodiment, resource orchestrator 108 can include hardware, software, or some combination thereof.

[0114] In at least one embodiment, as 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 can include a framework that supports software layer 830 and / or one or more applications 842 of application layer 840. In at least one embodiment, software 832 or applications 842 can include web-based service software or applications, respectively, such as services or applications provided by Amazon Web Services, Google Cloud, and Microsoft Azure. In at least one embodiment, the framework layer 820 can be, but is not limited to, a type of free and open-source software web application framework such as Apache Spark™ (hereinafter “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 822 can include a Spark driver to facilitate scheduling workloads supported by various layers of the data center 800. In at least one embodiment, the configuration manager 824 can be capable of configuring different layers, such as the software layer 830 and the framework layer 820 including Spark and the distributed file system 828 for supporting large-scale data processing. In at least one embodiment, the resource manager 826 can be capable of managing clustered or grouped computing resources mapped to or allocated for supporting the distributed file system 828 and the job scheduler 822. In at least one embodiment, the clustered or grouped computing resources can include grouped computing resources 814 on the data center infrastructure layer 810. In at least one embodiment, the resource manager 826 can coordinate with the resource orchestrator 812 to manage these mapped or allocated computing resources.

[0115] In at least one embodiment, software 832 included in the software layer 830 can include software used by at least a portion of the node C.R.s 816(1)-816(N), the grouped computing resources 814, and / or the distributed file system 828 of the framework layer 820. One or more types of software can include, but are not limited to, Internet web page search software, email virus scanning software, database software, and streaming video content software.

[0116] In at least one embodiment, one or more applications 842 included in application layer 840 can include one or more types of applications used by at least portions of node C.R.s 816(1)-816(N), grouped computing resources 814, and / or distributed file system 828 of framework layer 820. One or more types of applications can include, but are not limited to, any number and / or type of genomics applications, cognitive computing and machine learning applications including training or inferencing software, machine learning framework software (e.g., PyTorch, TensorFlow, Caffe, etc.), or other machine learning applications used in conjunction with one or more embodiments.

[0117] In at least one embodiment, any of configuration manager 824, resource manager 826, and resource orchestrator 812 can implement any number and type of self-modifying actions based on any amount and type of data acquired in any technically feasible fashion. In at least one embodiment, self-modifying actions can alleviate data center operators of data center 800 from making possibly poor configuration decisions and can avoid underutilized and / or poorly performing portions of a data center.

[0118] In at least one embodiment, data center 800 can 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 in accordance with one or more embodiments described herein. For example, in at least one embodiment, a machine learning model can be trained by computing weight parameters according to a neural network architecture using software and computing resources described above with respect to data center 800. In at least one embodiment, using weight parameters computed by one or more training techniques described herein, a trained machine learning model corresponding to one or more neural networks can be used to infer or predict information using resources described above with respect to data center 800.

[0119] In at least one embodiment, a data center can use CPUs, application specific integrated circuits (ASICs), GPUs, FPGAs, or other hardware to perform training and / or inferencing using resources described above. Moreover, one or more software and / or hardware resources described above can be configured as a service to allow users to train or perform information inference such as image recognition, speech recognition, or other artificial intelligence services.

[0120] Inference and / or training logic 715 are used to perform inferencing and / or training operations associated with one or more embodiments. Inference and / or training logic 715 are used to perform inferencing and / or training operations associated with one or more embodiments described herein using one or more examples of a neural network, of a support Figure 7A and / or Figure 7BDetails regarding inference and / or training logic 715 are provided. In at least one embodiment, inference and / or training logic 715 can be used in a system that uses neural network training operations, neural network functions and / or architectures, or neural network use cases described herein to infer or predict operations based, at least in part, on weight parameters calculated using neural network training operations, neural network functions and / or architectures, or neural network use cases described herein. Figure 8

[0121] Such components can be used to perform unsupervised domain transfer for one or more neural networks, which can include a feature extraction network and a classifier network.

[0122] Computer system

[0123] Figure 9 is a block diagram illustrating an example computer system, which can be a system with interconnected devices and components, a system on a chip (SOC), or some combination thereof formed with a processor that can include execution units to execute an instruction, according to at least one embodiment. In at least one embodiment, consistent with the present disclosure, for example, embodiments described herein, computer system 900 can include, without limitation, a component, such as processor 902, whose execution units include logic to perform an algorithm for process data. In at least one embodiment, computer system 900 can include a processor, such as a Pentium®, Core®, Xenon®, or Itanium® family of processors, XeonTM, or microprocessors, although the scope of the embodiments is not so limited. In at least one embodiment, consistent with the present disclosure, computer system 900 can perform versions of the WINDOWS operating systems from MICROSOFT CORPORATION of Redmond, Washington, although other operating systems, embedded software, and / or graphical user interfaces can be used. In at least one embodiment, consistent with the present disclosure, computer system 900 can include, without limitation, a system

[0124] ​Embodiments can 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, embedded applications can 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

[0125] In at least one embodiment, computer system 900 can include, but is not limited to, processor 902, which can include, but is not limited to, one or more execution units 908 to perform, e.g., machine learning model training and / or inferencing, in accordance with techniques described herein. In at least one embodiment, computer system 900 is a single processor desktop or server system, but in another embodiment, computer system 900 can be a multiprocessor system. In at least one embodiment, processor 902 can include, but is not limited to, a complex instruction set computer ("CISC") microprocessor, a reduced instruction set computing ("RISC") microprocessor, a very long instruction word ("VLIW") microprocessor, a processor implementing a combo of instruction sets, or any other processor device, such as a digital signal processor. In at least one embodiment, processor 902 can be coupled to a processor bus 910 that can transmit data signals between processor 902 and other components in computer system 900.

[0126] In at least one embodiment, processor 902 can include, but is not limited to, level 1 ("Ll") internal cache memory ("cache") 904. In at least one embodiment, processor 902 can have a single -level internal cache or multi-level internal cache. In at least one embodiment, cache memory can reside in the processor 902's external. Other embodiments can include a combination of internal and external caches based on specific implementation and requirements. In at least one embodiment, register file 906 can store different types of data within various registers including, but not limited to, integer registers, floating point registers, status registers, and instruction pointer registers.

[0127] In at least one embodiment, execution unit 908 includes, without limitation, logic to perform integer and floating-point operations, including bit- wide operations. In at least one embodiment, processor 902 can also include a microcode (“ucode”) read only memory (“ROM”) that stores microcode for certain macro instructions. In at least one embodiment, execution unit 908 can also include logic to handle a packed instruction set 909. In at least one embodiment, by including packed instruction set 909 in a general-purpose processor, many multimedia applications can be accelerated by using full width of data bus of processor 902. In one or more embodiments, by using full width of data bus of processor for one or more operations on packed data, many multimedia applications can be executed more efficiently and speedier.

[0128] In at least one embodiment, execution unit 908 can also be used in a microcontroller, embedded processor, graphics device, DSP, and other types of logic circuits. In at least one embodiment, computer system 900 can include, without limitation, memory 920. In at least one embodiment, memory 920 can be implemented as a Dynamic Random Access Memory (“DRAM”) device, a Static Random Access Memory (“SRAM”) device, a flash memory device, or other memory device. In at least one embodiment, memory 920 can store instruction(s) 919 and / or data 921 represented by data signals that can be executed by processor 902.

[0129] In at least one embodiment, a system logic chip can be coupled to processor bus 910 and memory 920. In at least one embodiment, system logic chip can include, without limitation, a memory controller hub (“MCH”) 916 and processor 902 can communicate with MCH 916 via processor bus 910. In at least one embodiment, MCH 916 can provide a high bandwidth memory path 918 to memory 920 for instruction and data storage and for storage of graphics commands, data, and textures. In at least one embodiment, MCH 916 can direct data signals between processor 902, memory 920, and other components in computer system 900, and can

[0130] In at least one embodiment, computer system 900 can use system I / O 922, which is a proprietary hub interface bus to couple MCH 916 to I / O controller hub (“ICH”) 930. In at least one embodiment, ICH 930 can provide a direct connection to some I / O devices and can indirect connect other devices via an input / output (I / O) bus.

[0131] In at least one embodiment, Figure 9 A system including interconnected hardware devices or “chips” is shown, while in other embodiments, Figure 9 An exemplary system on a chip (SoC) can be shown. In at least one embodiment, devices can be interconnected with a proprietary interconnect, a standardized 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.

[0132] The inference and / or training logic 715 is used to perform inference and / or training operations related to one or more embodiments. (The following is in conjunction with...) Figure 7A and / or Figure 7B Details are provided regarding the inference and / or training logic 715. In at least one embodiment, the inference and / or training logic 715 can be... Figure 9 Used in systems for reasoning or predicting operations based at least in part on weight parameters calculated using neural network training operations, neural network functions and / or architectures or neural network use cases described herein.

[0133] Such components can be used to perform unsupervised domain transfer for one or more neural networks, such as feature extraction networks and classifier networks.

[0134] Figure 10 This 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 laptop computer, tower server, rack server, blade server, desktop computer, tablet computer, mobile device, telephone, embedded computer, or any other suitable electronic device.

[0135] In at least one embodiment, system 1000 may include, but is not limited to, processor 1010 communicatively coupled to any suitable number or type of components, peripherals, modules, or devices. In at least one embodiment, processor 1010 uses a bus or interface coupling, such as an I2C bus, system management bus (“SMBus”), low pin count (LPC) bus, serial peripheral interface (“SPI”), high-definition audio (“HDA”) bus, serial advanced technology accessory (“SATA”) bus, universal serial bus (“USB”) (versions 1, 2, and 3), or universal asynchronous receiver / transmitter (“UART”) bus. In at least one embodiment, Figure 10 The system shown includes interconnected hardware devices or "chips," while in other embodiments, Figure 10 An exemplary system-on-a-chip (SoC) can be illustrated. In at least one embodiment, Figure 10 The device shown can be interconnected with proprietary interconnects, standardized interconnects (e.g., PCIe), or some combination thereof. In at least one embodiment, Figure 10 One or more components are interconnected using Computational Fast Link (CXL) interconnects.

[0136] In at least one embodiment, Figure 10The display 1024, touch screen 1025, touch pad 1030, near field communication unit ("NFC") 1045, sensor hub 1040, thermal sensor 1046, express chipset ("EC") 1035, trusted platform module ("TPM") 1038, BIOS / firmware / flash ("BIOS, FW Flash") 1022, DSP 1060, drive 1020 (e.g., solid state disk ("SSD") or hard disk drive ("HDD")), wireless local area network unit ("WLAN") 1050, Bluetooth unit 1052, wireless wide area network unit ("WWAN") 1056, Global Positioning System (GPS) 1055, camera ("USB3.0 camera") 1054 (e.g., USB3.0 camera), and / or low power double data rate ("LPDDR") memory unit ("LPDDR3") 1015 implemented in, for example, LPDDR3 standard, can each be implemented with any suitable structure, for example.

[0137] In at least one embodiment, other components can be communicatively coupled to processor 1010 by components described above. In at least one embodiment, accelerometer 1041, ambient light sensor ("ALS") 1042, compass 1043, and gyroscope 1044 can be communicatively coupled to sensor hub 1040. In at least one embodiment, thermal sensor 1039, fan 1037, keyboard 1036, and touch pad 1030 can be communicatively coupled to EC 1035. In at least one embodiment, speaker 1063, earpiece 1064, and microphone ("mic") 1065 can be communicatively coupled to audio unit ("audio codec and class D amplifier") 1062, which can in turn be communicatively coupled to DSP 1060. In at least one embodiment, audio unit 1062 can include, for example and without limitation, an audio coder / decoder ("codec") and a class D amplifier. In at least one embodiment, SIM card ("SIM") 1057 can be communicatively coupled to WWAN unit 1056. In at least one embodiment, components such as WLAN unit 1050 and Bluetooth unit 1052, as well as WWAN unit 1056, can be implemented as a next generation form factor (NGFF).

[0138] Inference and / or training logic 715 are used to perform inferencing and / or training operations associated with one or more embodiments. Details regarding inference and / or training logic 715 are described in more detail below. Figure 7A and / or Figure 7BDetails regarding inference and / or training logic 715 are provided. In at least one embodiment, inference and / or training logic 715 can be used in graphics double precision floating point unit 1000 for inferencing or predicting operations based, at least in part, on weight parameters calculated using neural network training operations, neural network functions and / or architectures, or neural network use cases described herein.

[0139] Such components can be used to perform unsupervised domain transfer for one or more neural networks, which can include a feature extraction network and a classifier network.

[0140] Figure 11 is a block diagram of a processing system in accordance with at least one embodiment. In at least one embodiment, system 1100 includes one or more processor(s) 1102 and one or more graphics processing unit(s) 1108, and can be a single processor desktop system, a multiprocessor workstation system, or a server system having many processor(s) 1102 or processor core(s) 1107. In at least one embodiment, system 1100 is a processing platform incorporated within a system on a chip (SoC) integrated circuit for use in mobile, handheld, or embedded devices.

[0141] In at least one embodiment, system 1100 can include or be incorporated within a server-based gaming platform, which for example, can be included in a gaming and media console. In at least one embodiment, system 1100 is a console that includes a game and media console with support for multiple game controllers and media controllers. In at least one embodiment, system 1100 can be integrated with one or more game controllers 1120. In at least one embodiment, system 1100 is a mobile phone, a smart phone, a phablet, a tablet, a laptop, a computer, a server, a gaming console, or other processing device.

[0142] In at least one embodiment, one or more processors 1102 each include one or more processor cores 1107 to process instructions which, 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 set 1109. In at least one embodiment, instruction set 1109 can facilitate complex instruction set computing (CISC), reduced instruction set computing (RISC), or computing via a very long instruction word (VLIW). In at least one embodiment, processor cores 1107 can each process a different instruction set 1109, which can include instructions to facilitate emulation of other instruction sets. In at least one embodiment, processor core 1107 can include other processing devices, such as a digital signal processor (DSP).

[0143] In at least one embodiment, processor 1102 includes cache memory 1104. In at least one embodiment, processor 1102 can have single-level or multi-level internal caches. In at least one embodiment, cache memory is shared among multiple components of processor 1102. In at least one embodiment, processor 1102 also uses an external cache (e.g., a three-level (L3) cache, or last level cache (LLC)) (not shown), which can be shared among processor cores 1107 using known cache coherence techniques. In at least one embodiment, additionally included in processor 1102 are register files 1106, which processor can include of different types of registers such as integer registers, floating point registers, status registers, and instruction pointer registers. In at least one embodiment, register files 1106 can include general registers or other registers.

[0144] In at least one embodiment, one or more processors 1102 are coupled with one or more interface buses 1110 for passing communication signals between processor 1102 and other components of system 1100. In at least one embodiment, one or more of interface buses 1110 can be versions of a Peripheral Component Interconnect (PCI) bus or PCI Express bus. In at least one embodiment, one or more of interface buses 1110 can be versions of an Accelerated Graphics Port (AGP) bus, or other bus types. In at least one embodiment, one or more of interface buses 1110 are direct media

[0145] In at least one embodiment, memory device 1120 can be a Dynamic Random Access Memory (DRAM) device, a Static Random Access Memory (SRAM) device, a flash memory device, or a content addressable memory device, among others. In at least one embodiment, memory device 1120 can be a Static Random Access Memory (SRAM) device, a flash memory device, or a content addressable memory device, among others. In at least one embodiment, memory device 1120 can be used as a system memory for processing system 1100, to store data 1122 and instructions 1121 for use when executing applications or processes by one or more processors 1102. In at least one embodiment, memory controller 1116 also couples with an optional external graphics processor 1112, which can communicate with one or more graphics processors 1108 in processor 1102 to perform graphics and media operations.

[0146] In at least one embodiment, platform controller hub 1130 enables peripherals coupled to bridge 1122 to interact with a processor and / or each other over high-speed I / O buses 1120 and 1110. In at least one embodiment, 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., solid-state drive (SSD), floppy drive, optical drive, etc.). In at least one embodiment, data storage device 1124 can communicate via a storage interface (e.g., SATA) or via a peripheral bus, such as a Peripheral Component Interconnect bus (e.g., PCI, PCI Express). In at least one embodiment, touch sensor 1125 can include a touch screen sensor, a pressure sensor, or a fingerprint sensor. In at least one embodiment, 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, firmware interface 1128 enables communication with system firmware, and can be, for example, a unified extensible firmware interface (UEFI). In at least one embodiment, network controller 1134 can enable a network connection to a wired network. In at least one embodiment, a high-performance network controller (not shown) couples with interface bus 1110. In at least one embodiment, audio controller 1146 is a multi-channel high definition audio controller. In at least one embodiment, processing system 1100 includes an optional legacy I / O controller 1140 for coupling legacy (e.g., Personal System 2 (PS / 2)) devices to system 1100. In at least one embodiment, platform controller hub 1130 can also connect to one or more Universal Serial Bus (USB) controllers 1142 connect input devices, such as keyboard and mouse 1143 combinations, camera 1144, or other USB input devices.

[0147] In at least one embodiment, memory controller 1116 and platform controller hub 1130 can 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 can be external to one or more processor(s) 1102. For example, in at least one embodiment, system 1100 can include an external memory controller 1116 and platform controller hub 1130, which can be configured as a memory controller hub and peripheral controller hub in a system-on-a-chip (SoC) that can communicate with processor(s) 1102.

[0148] Inference and / or training logic 715 are used to perform inferencing and / or training operations associated with one or more embodiments. Details regarding inference and / or training logic 715 are provided below in conjunction with FIGS. 7L and / or 7M. Figure 7A and / orFigure 7B Details regarding the inference and / or training logic 715 are provided. In at least one embodiment, some or all of inference and / or training logic 715 can be incorporated with graphics processor 1100. For example, in at least one embodiment, the training and / or inference techniques described herein can use one or more ALUs embodied in a graphics processor. Further, in at least one embodiment, the inference and / or training operations described herein can be accomplished with logic other than that illustrated in FIG. 7. Figure 7A or Figure 7B In at least one embodiment, weight parameters can be stored in on-chip or off-chip memory and / or registers (shown or not) that configure ALUs of a graphics processor to perform one or more machine learning algorithms, neural network architectures, use cases, or training techniques described herein.

[0149] Such components can be used to perform unsupervised domain transfer for one or more neural networks, which can include a feature extraction network and a classifier network.

[0150] Figure 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, processor 1200 can include additional cores, up to and including an additional core 1202N represented by a dashed lined in FIG. 12. 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 can also include access to one or more shared cache units 1206.

[0151] In at least one embodiment, internal cache units 1204A-1204N and shared cache unit 1206 represent a cache memory hierarchy within processor 1200. In at least one embodiment, cache memory units 1204A-1204N can include at least one level of cache memory such as level one (LI), level two (L2), level three (L3), level four (L4), or other levels of cache, within each processor core 1202A-1202N and shared level cache(s) 1206, where the highest level of cache memory prior to main memory is referred to as the LLC. In at least one embodiment, cache coherence logic maintains coherency for the various cache units 1206 and 1204A-1204N.

[0152] In at least one embodiment, processor 1200 also includes a set of one or more bus controller units 1216 and a system agent core 1210. In at least one embodiment, 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, system agent core 1210 provides management functionality for various processor components. In at least one embodiment, system agent core 1210 includes one or more integrated memory controllers 1214 to manage access to various external memory devices (not shown), including support for data bus protocols such as DDR SDRAM.

[0153] In at least one embodiment, one or more processor cores 1202A-1202N include support to run in multiple threads simultaneously. In at least one embodiment, system agent core 1210 includes components for coordination and operation of cores 1202A-1202N during multi-threaded processing. In at least one embodiment, system agent core 1210 can additionally include a power control unit (PCU), including logic and components to govern one or more power states of processor cores 1202A-1202N and graphics processor 1208.

[0154] In at least one embodiment, processor 1200 also includes graphics processor 1208, which can be configured to perform a graphics processing applications or a graphics processing- intensive applications. In at least one embodiment, graphics processor 1208 couples with shared cache unit 1206 and system agent core 1210, including one or more integrated memory controllers 1214. In at least one embodiment, system agent core 1210 also includes a display controller 1211 for driving one or more coupled displays to present graphics processor output to a

[0155] In at least one embodiment, ring based interconnect unit 1212 is used to couple the internal components of the processor 1200. In at least one embodiment, an alternative interconnect unit can be used, such as a point-to-point interconnect, a switched interconnect, or other technology. In at least one embodiment, graphics processor 1208 couples with the ring interconnect 1212 through I / O link 1213.

[0156] 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 module 1218 (e.g., eDRAM module). In at least one embodiment, each of processor cores 1202A-1202N and graphics processor 1208 uses embedded memory module 1218 as a shared last-level cache.

[0157] In at least one embodiment, processor cores 1202A-1202N are homogeneous cores executing a common instruction set architecture. In at least one embodiment, processor cores 1202A-1202N are heterogeneous in terms of instruction set architecture (ISA), with one or more processor cores 1202A-1202N executing a common instruction set, while one or more other processor cores 1202A-1202N execute a subset of the common instruction set or a different instruction set. In at least one embodiment, processor cores 1202A-1202N are heterogeneous in terms of microarchitecture, with one or more cores having relatively high power consumption coupled to one or more power cores having lower power consumption. In at least one embodiment, processor 1200 may be implemented on one or more chips or implemented as a SoC integrated circuit.

[0158] Inference and / or training logic 715 is used to perform inference and / or training operations associated with one or more embodiments. The following is combined with... Figure 7A and / or Figure 7B Details regarding the inference and / or training logic 715 are provided. 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... Figure 12 The graphics processor 1512, graphics core 1202A-1202N, or other components are used. Furthermore, in at least one embodiment, the inference and / or training operations described herein can use, except... Figure 7A Alternatively, it may be accomplished using logic other than that shown in 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), which configure the ALU of the graphics processor 1200 to execute one or more machine learning algorithms, neural network architectures, use cases, or training techniques described herein.

[0159] Such components can be used to perform unsupervised domain transfer for one or more neural networks, such as feature extraction networks and classifier networks.

[0160] Virtualization computing platform

[0161] Figure 13 This is an example data flow diagram of process 1300 for generating and deploying an image processing and inference pipeline according to at least one embodiment. In at least one embodiment, process 1300 may be deployed for use with imaging devices, processing devices, and / or other device types at one or more facilities 1302. Process 1300 may be executed within training system 1304 and / or deployment system 1306. In at least one embodiment, training system 1304 may be used to train, deploy, and implement machine learning models (e.g., neural networks, object detection algorithms, computer vision algorithms, etc.) for use in deployment system 1306. In at least one embodiment, deployment system 1306 may be configured to offload processing and computing resources in a distributed computing environment to reduce the infrastructure requirements of facility 1302. In at least one embodiment, one or more applications in the pipeline may use or invoke services of deployment system 1306 (e.g., inference, visualization, computation, AI, etc.) during application execution.

[0162] In at least one embodiment, some applications used in the advanced processing and inference pipeline may use machine learning models or other AI to perform one or more processing steps. In at least one embodiment, a machine learning model may be trained at facility 1302 using data 1308 (e.g., imaging data) generated at facility 1302 (and stored on one or more Picture Archiving and Communication System (PACS) servers at facility 1302), imaging or sequencing data 1308 from another or more facilities, or a combination thereof. In at least one embodiment, training system 1304 may be used to provide applications, services, and / or other resources to generate a deployable machine learning model for the work of deploying system 1306.

[0163] In at least one embodiment, the model registry 1324 may be supported by an object storage system that supports version control and object metadata. In at least one embodiment, it may be available from within a cloud platform via, for example, cloud storage (e.g., Figure 14 The system uses a cloud-compatible application programming interface (API) (1426) to access object storage. In at least one embodiment, machine learning models within the model registry 1324 can be uploaded, listed, modified, or deleted by the developer or partner 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 models with applications, enabling the models to be executed as part of the containerized instantiation of the application.

[0164] In at least one embodiment, training pipeline 1404 ( Figure 14) can include instances in which 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 imaging devices, sequencing devices, and / or other types of devices can be received. In at least one embodiment, once imaging data 1308 is received, AI-assisted annotation 1310 can be used to help generate annotations corresponding to imaging data 1308 to be used as ground truth data for a machine learning model. In at least one embodiment, AI-assisted annotation 1310 can include one or more machine learning models (e.g., a convolutional neural network (CNN)) that can be trained to generate annotations corresponding to certain types of imaging data 1308 (e.g., from certain devices). In at least one embodiment, AI-assisted annotation 1310 can then be used directly, or can be adjusted or fine-tuned using annotation tools to generate ground truth data. In at least one embodiment, AI-assisted annotation 1310, labeled clinical data 1312, or a combination thereof can be used as ground truth data to train a machine learning model. In at least one embodiment, a trained machine learning model can be referred to as output model 1316, and can be used by deployment system 1306, as described herein.

[0165] In at least one embodiment, training pipeline 1404( Figure 14) can include situations in which facility 1302 needs a machine learning model for performing one or more processing tasks for one or more applications in deployment system 1306, but facility 1302 can not currently have such a machine learning model (or can not have a model that is optimized, efficient, or effective for this purpose). In at least one embodiment, an existing machine learning model can be selected from model registry 1324. In at least one embodiment, model registry 1324 can include machine learning models that are trained to perform a variety of different inferencing tasks on imaging data. In at least one embodiment, the machine learning models in model registry 1324 can have been trained on imaging data from different facilities (e.g., facilities located remotely from facility 1302). In at least one embodiment, a machine learning model can have been trained on imaging data from one location, two locations, or any number of locations. In at least one embodiment, when training on imaging data from a particular location, the training can be performed at that location, or at least in a manner that protects the confidentiality of the imaging data or limits migration of the imaging data off premises. In at least one embodiment, once a model is trained, or partially trained, at a location, the machine learning model can be added to 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 model registry 1324. In at least one embodiment, a machine learning model can then be selected from model registry 1324 (and referred to as output model 1316), and can be used in deployment system 1306 to perform one or more processing tasks for one or more applications of the deployment system.

[0166] In at least one embodiment, training pipeline 1404( Figure 14In at least one embodiment, scenario can include facility 1302 that requires a machine learning model for performing one or more processing tasks for deploying one or more applications in deployment system 1306, but facility 1302 can not currently have such a machine learning model (or can not have an optimized, efficient, or effective model). In at least one embodiment, due to population differences, robustness of training data used to train a machine learning model, diversity of training data anomalies, and / or other issues with training data, a machine learning model selected from model registry 1324 can not be fine-tuned or optimized for imaging data 1308 generated at facility 1302. In at least one embodiment, AI assisted annotation 1310 can be used to help generate annotations corresponding to imaging data 1308 for use as ground truth data to train or update a machine learning model. In at least one embodiment, labeled clinical data 1312 can be used as ground truth data to train a machine learning model. In at least one embodiment, retraining or updating a machine learning model can be referred to as model training 1314. In at least one embodiment, model training 1314 (e.g., AI assisted annotation 1310, labeled clinical data 1312, or a combination thereof) can be used as ground truth data to retrain or update a machine learning model. In at least one embodiment, a trained machine learning model can be referred to as output model 1316 and can be used by deployment system 1306, as described herein.

[0167] In at least one embodiment, deployment system 1306 can include software 1318, services 1320, hardware 1322, and / or other components, features, and functionality. In at least one embodiment, deployment system 1306 can include a software “stack” such that software 1318 can be built on top of services 1320, and can use services 1320 to perform some or all processing tasks, and services 1320 and software 1318 can be built on top of hardware 1322 and use hardware 1322 to perform processing, storage, and / or other computing tasks of deployment system. In at least one embodiment, software 1318 can include any number of different containers, where each container can execute an instantiation of an application. In at least one embodiment, each application can perform one or more processing tasks in a high-level processing and inference pipeline (e.g., inference, object detection, feature detection, segmentation, image enhancement, calibration, etc.). In at least one embodiment, a high-level processing and inference pipeline can be defined based on a selection of different containers desired or required to process imaging data 1308 (e.g., to convert output back to a usable data type. In at least one embodiment, a combination of containers within software 1318 (e.g., which make up a pipeline) can be referred to as a virtual instrument (as described in greater detail herein), and a virtual instrument can utilize services 1320 and hardware 1322 to perform some or all processing tasks of applications instantiated in containers.

[0168] In at least one embodiment, a 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 deployment system 1306). In at least one embodiment, 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, data can be pre-processed as part of a data processing pipeline to prepare data for processing by one or more applications. In at least one embodiment, post-processing can be performed on output of one or more inference tasks or other processing tasks of a pipeline to prepare output data for a next application and / or to prepare output data for transmission and / or use by a user (e.g., in response to an inference request). In at least one embodiment, inference tasks can be performed by one or more machine learning models, such as trained or deployed neural networks, which can include output models 1316 of training system 1304.

[0169] In at least one embodiment, the tasks of the data processing pipeline can 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 can be published to a private (e.g., limited access) area of ​​a container registry (described in more detail herein), and trained or deployed models can be stored in a 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) can be used in the container registry, and once a user selects an image from the container registry for deployment in the pipeline, that image can be used to generate containers for instantiation of the application for use by the user's system.

[0170] In at least one embodiment, a developer (e.g., a software developer, clinician, physician, etc.) can develop, publish, and store an application (e.g., as a container) for performing image processing and / or inference on provided data. In at least one embodiment, a software development kit (SDK) associated with the system can be used to perform development, publication, and / or storage (e.g., to ensure that the developed application and / or container conforms to or is compatible with the system). In at least one embodiment, the developed application can be tested locally using the SDK (e.g., at a first facility, testing data from a first facility), the SDK serving as a system (e.g.,...). Figure 14 System 1400 may support at least some services 1320. In at least one embodiment, since DICOM objects may contain one to hundreds of images or other data types, and due to variations in the data, the developer may be responsible for managing (e.g., setting up constructs for preprocessing built into the application, etc.) the extraction and preparation of incoming data. In at least one embodiment, once verified by system 1400 (e.g., for accuracy), the application becomes available in the container registry for user selection and / or implementation to perform one or more processing tasks on data at the user's facility (e.g., a second facility).

[0171] In at least one embodiment, the developer can then share the application or container over a network for the system (e.g., Figure 14of the system 1400) to access and use. In at least one embodiment, completed and validated applications or containers can be stored in a container registry, and related machine learning models can be stored in a model registry 1324. In at least one embodiment, a requesting entity (which provides an inference or image processing request) can browse the container registry and / or model registry 1324 to obtain applications, containers, datasets, machine learning models, etc., select a desired combination of elements to include in a data processing pipeline, and submit an image processing request. In at least one embodiment, a request can include input data necessary to perform the request (and, in some examples, data related to a patient), and / or can include a selection of applications and / or machine learning models to be executed in processing the request. In at least one embodiment, a request can then be passed to one or more components of the deployment system 1306 (e.g., a cloud) to perform processing of the data processing pipeline. In at least one embodiment, processing by the deployment system 1306 can include referencing elements (e.g., applications, containers, models, etc.) selected from the container registry and / or model registry 1324. In at least one embodiment, once results are generated through the pipeline, the results can be returned to a user for review (e.g., for review in a viewing application suite executed on a local, on-premises workstation or terminal).

[0172] In at least one embodiment, to help process or execute applications or containers in a pipeline, services 1320 can be utilized. In at least one embodiment, services 1320 can include computing services, artificial intelligence (Al) services, visualization services, and / or other service types. In at least one embodiment, services 1320 can provide functionality that is common to one or more applications in software 1318, and thus functionality can be abstracted as a service that can be called or utilized by applications. In at least one embodiment, functionality provided by services 1320 can run dynamically and more efficiently, while also allowing applications to process data in parallel (e.g., using Figure 14the shared services 1320 to provide the same functionality, but rather the services 1320 can be shared among and between various applications. In at least one embodiment, by way of non-limiting example, services can include inference servers or engines that can be used to perform detection or segmentation tasks. In at least one embodiment, a model training service can be included, which can provide machine learning model training and / or retraining capabilities. In at least one embodiment, a data augmentation service can be further included, which can provide GPU-accelerated data (e.g., DICOM, RIS, CIS, REST-compliant, RPC, raw, etc.) extraction, resizing, scaling, and / or other augmentations. In at least one embodiment, a visualization service can be used, which can add image rendering effects (e.g., ray tracing, rasterization, de-noising, sharpening, etc.) to add realism to two-dimensional (2D) and / or three-dimensional (3D) models. In at least one embodiment, a virtual instrument service can be included, which provides beamforming, segmentation, inference, imaging, and / or support to other applications within a pipeline of a virtual instrument.

[0173] In at least one embodiment, where the services 1320 include an AI service (e.g., an inference service), as part of execution of an application, one or more machine learning models can be executed by invoking (e.g., as an API call) the inference service (e.g., 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 invoke 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, software 1318 implementing an advanced processing and inference pipeline, which includes a segmentation application and an anomaly detection application, can be pipelined, as each application can invoke the same inference service to perform one or more inference tasks. In at least one embodiment, hardware 1322 can include GPUs, CPUs, graphics cards, AI / deep learning systems (e.g., AI supercomputers such as NVIDIA’s DGX), cloud platforms, or a combination thereof.

[0174] In at least one embodiment, different types of hardware 1322 can be used to provide efficient, specially-built support for software 1318 and services 1320 in deployment system 1306. In at least one embodiment, GPU processing can be implemented for local processing within an AI / deep learning system, in a cloud system, and / or in other processing components of deployment system 1306 (e.g., at facility 1302) to improve efficiency, accuracy, and performance of image processing and generation. In at least one embodiment, software 1318 and / or services 1320 can be optimized for GPU processing, by way of non-limiting example with respect to deep learning, machine learning, and / or high performance computing. In at least one embodiment, at least some of computing environments of deployment system 1306 and / or training system 1304 can be executed in a data center, one or more supercomputer or high performance computer systems with GPU-optimized software (e.g., a combination of hardware and software of NVIDIA DGX systems). In at least one embodiment, hardware 1322 can include any number of GPUs that can be called upon to perform data processing in parallel, as described herein. In at least one embodiment, a cloud platform can also include GPU processing for GPU-optimized execution of deep learning tasks, 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 systems) can be used as a hardware abstraction and scaling platform to execute a cloud platform (e.g., NVIDIA’s NGC), in at least one embodiment. In at least one embodiment, a cloud platform can integrate an application container clustering system or orchestration system (e.g., KUBERNETES) across multiple GPUs to enable seamless scaling and load balancing.

[0175] Figure 14 is a system diagram of an example system 1400 for generating and deploying imaging deployment pipelines, in accordance with at least one embodiment. In at least one embodiment, system 1400 can be used to implement process 1300 of FIG. 13 and / or other processes, including advanced processing and inference pipelines. In at least one embodiment, system 1400 can include training system 1304 and deployment system 1306. In at least one embodiment, training system 1304 and deployment system 1306 can be implemented using software 1318, services 1320, and / or hardware 1322, as described herein.

[0176] 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 medical 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 limited to authorized users by instituting security measures or protocols. In at least one embodiment, security protocols can include network tokens that can be signed by an authentication (e.g., AuthN, AuthZ, Gluecon, etc.) service and can carry appropriate authorization. In at least one embodiment, APIs (described herein) of a virtual instrument or other instances of system 1400 can be limited to a set of public IPs that have been vetted or authorized for interaction.

[0177] In at least one embodiment, various components of system 1400 can communicate information between each other using any of a plurality 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, communication between facilities and components of system 1400 (e.g., for sending inference requests, for receiving results of inference requests, etc.) can be communicated through one or more data buses, wireless data protocols (Wi-Fi), wired data protocols (e.g., Ethernet), etc.

[0178] In at least one embodiment, similar to training pipelines 1302 described herein with respect to Figure 13 In at least one embodiment, 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), where deployment system 1306 will use the one or more machine learning models in deployment pipeline 1410. In at least one embodiment, as a result of training pipeline 1404, an output model 1316 can be generated. In at least one embodiment, 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 deployment system 1306. In at least one embodiment, a training pipeline 1404 similar to the training pipeline 1302 described with respect to Figure 13 the first example described with respect to Figure 13 the second example described with respect to Figure 13The training pipeline 1404 of the third example described can be used for the third machine learning model. In at least one embodiment, any combination of tasks within the training system 1304 can be used depending on the requirements of each respective machine learning model. In at least one embodiment, one or more machine learning models can already be trained and ready for deployment, so the training system 1304 can not perform any processing on the machine learning model and the one or more machine learning models can be implemented by the deployment system 1306.

[0179] In at least one embodiment, the output model 1316 and / or the pre-trained model 1406 can include any type of machine learning model, depending on implementation or embodiment. In at least one embodiment and without limitation thereto, machine learning models used by the system 1400 can include using linear regression, logistic regression, decision trees, support vector machines (SVM), Naive Bayes, k- nearest neighbors (Knn), k-means clustering, random forests, dimensionality reduction algorithms, gradient boosting algorithms, neural networks (e.g., autoencoders, convolutional, recurrent, perceptrons, long / short-term memory (LSTM), Hopfield, Boltzmann, deep belief, deconvolutional, generative adversarial, liquid state machines, etc.), and / or other types of machine learning models.

[0180] In at least one embodiment, the training pipeline 1404 can include AI-assisted annotation, as described herein with respect to at least Figure 15BIn at least one embodiment, labeled clinical data 1312 (e.g., traditional annotations) can be generated by any number of techniques. In at least one embodiment, labels or other annotations can be generated in a drawing program (e.g., 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 can be hand drawn, in some examples. In at least one embodiment, ground truth data can be synthetically generated (e.g., from computer models or renderings), realistically generated (e.g., designed and generated from real world data), automatically generated by a machine (e.g., using feature analysis and learning to extract features from data and then generate labels), manually annotated (e.g., by a marker or annotation specialist defining locations of labels), and / or combinations thereof. In at least one embodiment, for each instance of imaging data 1308 (or other data types used by a machine learning model), there can be corresponding ground truth data generated by training system 1304. In at least one embodiment, AI-assisted annotation can be performed as part of deployment pipeline 1410; in addition to or instead of AI-assisted annotation included in training pipeline 1404. In at least one embodiment, system 1400 can include a multi-tiered platform that can include a software tier of diagnostic applications (or other application types) (e.g., software 1318) that can perform one or more medical imaging and diagnostic functions. In at least one embodiment, system 1400 can be communicatively coupled to (e.g., via encrypted links) a PACS server network of one or more facilities. In at least one embodiment, system 1400 can be configured to access and reference data from a PACS server to perform operations such as training machine learning models, deploying machine learning models, image processing, inferencing, and / or other operations.

[0181] In at least one embodiment, a software tier can be implemented as a secure, encrypted, and / or authenticated API through which applications or containers can be invoked (e.g., called) from an external environment (e.g., facility 1302). In at least one embodiment, an application can then call or execute one or more services 1320 to perform computing, AI, or visualization tasks associated with respective applications, and software 1318 and / or services 1320 can utilize hardware 1322 to perform processing tasks in an efficient and effective manner.

[0182] In at least one embodiment, deployment system 1306 can execute deployment pipelines 1410. In at least one embodiment, deployment pipelines 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 annotation - generated by imaging devices, sequencing devices, genomics devices, etc., as described above. In at least one embodiment, a deployment pipeline 1410 for an individual device can be referred to as a virtual instrument for a device (e.g., a virtual ultrasound instrument, a virtual CT scan instrument, a virtual sequencing instrument, etc.), as described herein. In at least one embodiment, there can be more than one deployment pipeline 1410 for a single device, depending on information desired from data generated by a device. In at least one embodiment, where anomalies are desired to be detected from MRI machines, there can be a first deployment pipeline 1410, and where image enhancement is desired from output of MRI machines, there can be a second deployment pipeline 1410.

[0183] In at least one embodiment, image generation applications can include processing tasks that include use of machine learning models. In at least one embodiment, a user can wish to use their own machine learning model, or select a machine learning model from model registry 1324. In at least one embodiment, a user can implement their own machine learning model or select a machine learning model for inclusion in an application that performs a processing task. In at least one embodiment, applications can be selectable and customizable, and by defining a construction of an application, deployment and implementation of an application for a particular user is presented as a more seamless user experience. In at least one embodiment, by leveraging other features of system 1400 (e.g., services 1320 and hardware 1322), deployment pipelines 1410 can be more user friendly, provide easier integration, and produce more accurate, efficient, and timely results.

[0184] In at least one embodiment, deployment system 1306 can include a user interface 1414 (e.g., a graphical user interface, a web interface, etc.) that can be used to select applications to include in deployment pipelines 1410, arrange applications, modify or change applications or parameters or constructions thereof, use and interact with deployment pipelines 1410 during setup and / or deployment, and / or otherwise interact with deployment system 1306. In at least one embodiment, although not shown with respect to training system 1304, user interface 1414 (or a different user interface) can be used to select models for use in deployment system 1306, for selecting models for training or retraining in training system 1304, and / or for otherwise interacting with training system 1304.

[0185] In at least one embodiment, in addition to application orchestration system 1428, pipeline manager 1412 can also be used to manage interactions between applications or containers of deployment pipeline 1410 and services 1320 and / or hardware 1322. In at least one embodiment, pipeline manager 1412 can be configured to facilitate interactions from application to application, from application to service 1320, and / or from application or service to hardware 1322. In at least one embodiment, although illustrated as included in software 1318, this is not intended to be limiting, and in some examples (e.g., as shown), pipeline manager 1412 can be included in services 1320. In at least one embodiment, application orchestration system 1428 (e.g., Kubernetes, DOCKER, etc.) can include a container orchestration system that can group applications into containers as logical units for orchestration, management, scaling, and deployment. In at least one embodiment, by associating applications (e.g., rebuilt applications, split applications, etc.) from deployment pipeline 1410 with individual containers, each application can execute in a self-contained environment (e.g., at kernel level) to improve speed and efficiency. Figure 14

[0186] ​In at least one embodiment, each application and / or container (or image thereof) can be separately developed, modified, and deployed (e.g., a first user or developer can develop, modify, and deploy a first application, a second user or developer can develop, modify, and deploy a second application separate from first user or developer), which can allow for focus and attention to tasks of a single application and / or container without being impeded by tasks of another application or container. In at least one embodiment, pipeline manager 1412 and application coordination system 1428 can facilitate communication and collaboration between different containers or applications. In at least one embodiment, as long as intended inputs and / or outputs of each container or application are known to system (e.g., based on construction of application or container), application coordination system 1428 and / or pipeline manager 1412 can facilitate communication between and among each application or container and sharing of resources. In at least one embodiment, as one or more applications or containers in deployment pipeline 1410 can share same services and resources, application coordination system 1428 can coordinate, load balance, and determine sharing of services or resources between and among various applications or containers. In at least one embodiment, a scheduler can be used to track resource needs of applications or containers, current or planned use of these resources, and resource availability. Accordingly, in at least one embodiment, a scheduler can allocate resources to different applications and among and between applications, taking into account needs and availability of system. In some examples, a scheduler (and / or other components of application coordination system 1428) can determine resource availability and distribution based on constraints imposed on system (e.g., user constraints), such as quality of service (QoS), urgency of data output (e.g., to determine whether to perform real-time processing or delayed processing), etc.

[0187] In at least one embodiment, services 1320 utilized by and shared by applications or containers in deployment system 1306 can include compute services 1416, AI services 1418, visualization services 1420, and / or other service types. In at least one embodiment, applications can invoke (e.g., execute) one or more services 1320 to perform processing operations for applications. In at least one embodiment, applications can utilize compute services 1416 to perform supercomputing or other high performance computing (HPC) tasks. In at least one embodiment, one or more compute services 1416 can be utilized to perform parallel processing (e.g., using 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, parallel computing platform 1430 (e.g., NVIDIA’s CUDA) can enable general purpose computing on GPUs (GPGPU) (e.g., GPU 1422). In at least one embodiment, software layers of parallel computing platform 1430 can provide access to virtual instruction sets and parallel computing elements of GPUs to execute compute kernels. In at least one embodiment, parallel computing platform 1430 can include memory, and in some embodiments, can share memory 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 can be generated for multiple containers and / or multiple processes within a container to use same data from a shared memory segment of parallel computing platform 1430 (e.g., where multiple different stages of an application or applications are processing same information). In at least one embodiment, rather than copying data and moving data to different locations in memory (e.g., read / write operations), same data in same location in memory can be used for any number of processing tasks (e.g., at same time, at different times, etc.). In at least one embodiment, as data is used to generate new data as a result of processing, this information of new location of data can be stored and shared between various applications. In at least one embodiment, location of data, as well as location of updated or modified data, can be part of a definition of how to understand a payload in a container.

[0188] In at least one embodiment, AI services 1418 can be utilized to perform inferencing services for executing machine learning models associated with applications (e.g., task is to perform one or more processing tasks for an application). In at least one embodiment, AI services 1418 can utilize AI system 1424 to execute machine learning models (e.g., neural networks such as CNNs) for segmentation, reconstruction, object detection, feature detection, classification, and / or other inferencing tasks. In at least one embodiment, an application of deployment pipeline 1410 can use one or more output models 1316 from training system 1304 and / or other models of an application to perform inferencing on imaging data. In at least one embodiment, two or more examples of inferencing can be available using an application coordination system 1428 (e.g., a scheduler). In at least one embodiment, a first category can include a high priority / low latency path, which can implement a higher service level agreement, such as for performing inferencing on urgent requests in emergency situations, or for radiologists during a diagnosis process. In at least one embodiment, a second category can include a standard priority path, which can be used for requests that can not be urgent or can have analysis performed at a later time. In at least one embodiment, application coordination system 1428 can allocate resources (e.g., services 1320 and / or hardware 1322) based on a priority path for different inferencing tasks of AI services 1418.

[0189] In at least one embodiment, shared storage can be installed to AI services 1418 in system 1400. In at least one embodiment, shared storage can operate as a cache (or other storage device type) and can 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 deployment system 1306 can receive the request and can select one or more instances (e.g., for best fit, for load balancing, etc.) to process the request. In at least one embodiment, to process the request, the request can be input into a database, a machine learning model can be located from model registry 1324 if not already in cache, a validation step can ensure that appropriate machine learning model is loaded into cache (e.g., shared storage), and / or a copy of the model can be saved to cache. In at least one embodiment, if an application has not already been running or there are not enough instances of an application, a scheduler (e.g., of pipeline manager 1412) can be used to start the application referenced in the request. In at least one embodiment, if an inference server has not already been started to execute the model, an inference server can be started. Each model can start any number of inference servers. In at least one embodiment, in a pull model of clustering inference servers, a model can be cached whenever load balancing is favorable. In at least one embodiment, inference servers can be statically loaded into respective distributed servers.

[0190] In at least one embodiment, inference can be performed using inference servers running in containers. In at least one embodiment, an instance of an inference server can be associated with a model (and optionally multiple versions of a model). In at least one embodiment, if an instance of an inference server does not exist when a request to perform inference on a model is received, a new instance can be loaded. In at least one embodiment, when an inference server is started, a model can be passed to the inference server so that the same container can be used to service different models as long as the inference server is run as a different instance.

[0191] In at least one embodiment, during application execution, an inference request for a given application can be received and a container (e.g., an instance hosting an inference server) can be loaded (if not already loaded) and a launcher can be invoked. In at least one embodiment, pre-processing logic in a container can load, decode, and / or perform any additional pre-processing on incoming data (e.g., using CPU and / or GPU). In at least one embodiment, once data is ready for inference, a container can infer on data as needed. In at least one embodiment, this can include a single inference call on one image (e.g., a hand X-ray), or can require inference on hundreds of images (e.g., a chest CT). In at least one embodiment, an application can summarize results before completion, which can include, without limitation, a single confidence score, a pixel-level segmentation, a voxel-level segmentation, generating a visualization, or generating text to summarize results. In at least one embodiment, different priorities can be assigned for different models or applications. For example, some models can have real-time (TAT less than 1 minute) priority, while other models can have lower priority (e.g., TAT less than 10 minutes). In at least one embodiment, model execution time can be measured from a requesting authority or entity, and can include cooperative network traversal time as well as execution time of an inference service.

[0192] In at least one embodiment, transfer of requests between service 1320 and inference applications can be hidden behind a software development kit (SDK), and robust transfer 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 SDK will pull requests from queues and provide requests to applications. In at least one embodiment, a name of a queue can be provided in an environment from which SDK will pick up queues. In at least one embodiment, asynchronous communication through queues can be useful because it can allow any instance of an application to pick up work when it is available. Results can be transferred back through queues to ensure no data loss. In at least one embodiment, queues can also provide an ability to split work, because highest priority work can go into a queue that connects to most instances of an application, while lowest priority work can go into a queue that connects to a single instance that processes tasks in order of receipt. In at least one embodiment, an application can run on GPU-accelerated instances that are spawned in cloud 1426, and inference service can perform inference on GPUs.

[0193] In at least one embodiment, visualization service 1420 can be utilized to generate visualizations for viewing application and / or deployment pipeline 1410 output. In at least one embodiment, visualization service 1420 can utilize GPU 1422 to generate visualizations. In at least one embodiment, visualization service 1420 can implement rendering effects such as ray tracing to generate higher quality visualizations. In at least one embodiment, visualizations can include, without limitation, 2D image rendering, 3D volume rendering, 3D volume reconstruction, 2D tomographic slices, virtual reality displays, augmented reality displays, etc. In at least one embodiment, virtualization environment can be used to generate virtual interactive displays or environments (e.g., virtual environments) for interaction by system users (e.g., doctors, nurses, radiologists, etc.). In at least one embodiment, visualization service 1420 can include internal visualizers, movie and / or other rendering or image processing capabilities or functionality (e.g., ray tracing, rasterization, internal optics, etc.).

[0194] In at least one embodiment, hardware 1322 can include GPU 1422, AI system 1424, cloud 1426, and / or any other hardware used to execute training system 1304 and / or deployment system 1306. In at least one embodiment, GPU 1422 (e.g., NVIDIA’s TESLA and / or QUADRO GPUs) can include any number of GPUs that can be used to perform processing tasks for any features or functionality of computing service 1416, AI service 1418, visualization service 1420, other services, and / or software 1318. For example, for AI service 1418, GPU 1422 can be used to perform pre-processing on imaging data (or other data types used by machine learning models), post-processing on outputs of machine learning models, and / or perform inferencing (e.g., to execute machine learning models). In at least one embodiment, cloud 1426, AI system 1424, and / or other components of system 1400 can use GPU 1422. In at least one embodiment, cloud 1426 can include a GPU-optimized platform for deep learning tasks. In at least one embodiment, AI system 1424 can use GPUs, and one or more AI systems 1424 can be used to perform cloud 1426 (or at least portions of tasks that are deep learning or inferencing). Likewise, although hardware 1322 is shown as discrete components, this is not intended to be limiting, and any component of hardware 1322 can be combined with, or utilized by, any other component of hardware 1322.

[0195] In at least one embodiment, AI system 1424 can include a purpose-built computing system (e.g., a supercomputer or HPC) configured for inferencing, deep learning, machine learning, and / or other artificial intelligence tasks. In at least one embodiment, AI system 1424 (e.g., NVIDIA’s DGX) can include software (e.g., a software stack) that can use multiple GPUs 1422 to perform split-GPU optimizations in addition to CPUs, RAM, storage, and / or other components, features, or functions. In at least one embodiment, one or more AI systems 1424 can be implemented in cloud 1426 (e.g., in a data center) to perform some or all AI-based processing tasks of system 1400.

[0196] In at least one embodiment, cloud 1426 can include a GPU-accelerated infrastructure (e.g., NVIDIA’s NGC) that can provide a GPU-optimized platform for performing processing tasks of system 1400. In at least one embodiment, cloud 1426 can include AI system 1424 for performing 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 can integrate with application orchestration system 1428 to enable seamless scaling and load balancing between and among applications and services 1320 utilizing multiple GPUs. In at least one embodiment, cloud 1426 can be responsible for performing at least some services 1320 of system 1400, including compute services 1416, AI services 1418, and / or visualization services 1420, as described herein. In at least one embodiment, cloud 1426 can perform batched inference (e.g., perform NVIDIA’s TENSORRT), provide an accelerated parallel computing API and platform 1430 (e.g., NVIDIA’s CUDA), perform application orchestration system 1428 (e.g., KUBERNETES), provide a graphics rendering API and platform (e.g., for ray tracing, 2D graphics, 3D graphics, and / or other rendering techniques to produce higher quality cinematic effects), and / or can provide other functionality for system 1400.

[0197] Figure 15A A dataflow graph for process 1500 for training, retraining, or updating a machine learning model is shown, in accordance with at least one embodiment. In at least one embodiment, process 1500 can be performed using, as a non-limiting example, NVIDIA’s Figure 14The process 1500 can be performed by the system 1400. In at least one embodiment, the process 1500 can utilize the services 1320 and / or hardware 1322 of the system 1400, as described herein. In at least one embodiment, the refined model 1512 generated by the process 1500 can be executed by the deployment system 1306 for one or more containerized applications in the deployment pipeline 1410.

[0198] In at least one embodiment, the model training 1314 can include retraining or updating the initial model 1504 (e.g., a pre-trained model) using new training data (e.g., new input data (such as the customer dataset 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 layers of the initial model 1504 can be reset or deleted, and / or replaced with updated or new output or loss layers. In at least one embodiment, the initial model 1504 can have previously fine-tuned parameters (e.g., weights and / or biases) that are retained from previous training, so the training or retraining 1314 can not take as long or require as much processing as training a model from scratch. In at least one embodiment, during the model training 1314, by resetting or replacing the output or loss layers of the initial model 1504, the parameters of the new dataset can be updated and re-tuned as predictions are generated on the new customer dataset 1506 (e.g., image data 1308) based on loss calculations associated with the accuracy of the output or loss layers. Figure 13 In at least one embodiment, the pre-trained model 1406 can be stored in a data store or registry (e.g., the data store 1302) and accessed by the model training 1314 to generate the refined model 1512.

[0199] In at least one embodiment, the pre-trained model 1406 can be stored in a data store or registry (e.g., the data store 1302) and accessed by the model training 1314 to generate the refined model 1512. Figure 13model registry 1324). In at least one embodiment, pre-trained models 1406 can have been trained, at least in part, at one or more facilities other than the facility executing process 1500. In at least one embodiment, to protect the privacy and rights of patients, subjects, or customers of different facilities, pre-trained models 1406 can have been trained locally using locally generated customer or patient data. In at least one embodiment, pre-trained models 1406 can be trained using cloud 1426 and / or other hardware 1322, but confidential, privacy protected patient data can not be transferred to, used by, or accessed by any component of cloud 1426 (or other non-local hardware). In at least one embodiment, if pre-trained models 1406 are trained using patient data from more than one facility, pre-trained models 1406 can have been individually trained for each facility before training on patient or customer data from another facility. In at least one embodiment, 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, for example, in cases where customer or patient data has been de-identified (e.g., by waiver, for experimental use, etc.), or where customer or patient data is included in a public dataset.

[0200] In at least one embodiment, when selecting an application to use in deployment pipeline 1410, a user can also select a machine learning model for use with the particular application. In at least one embodiment, a user can not have a model to use, so the user can select a pre-trained model 1406 to use with the application. In at least one embodiment, pre-trained models 1406 can not be optimized for generating accurate results on a customer dataset 1506 of a user’s facility (e.g., based on patient diversity, demographics, types of medical imaging devices used, etc.). In at least one embodiment, pre-trained models 1406 can be updated, retrained, and / or fine-tuned for use at individual facilities before being deployed into deployment pipeline 1410 for use with one or more applications.

[0201] In at least one embodiment, a user can select a pre-trained model 1406 to update, retrain, and / or fine-tune, and the pre-trained model 1406 can be referred to as an initial model 1504 for training system 1304 in process 1500. In at least one embodiment, a customer dataset 1506 (e.g., imaging data, genomic data, sequencing data, or other data types generated by equipment at a facility) can be used to perform model training 1314 (which can include, without limitation, transfer learning) on initial model 1504 to generate a refined model 1512. In at least one embodiment, ground truth data corresponding to customer dataset 1506 can be generated by training system 1304. In at least one embodiment, ground truth data can be generated at least in part by a clinician, scientist, physician, practitioner at a facility (e.g., as labeled clinical data 1312 in FIG. 13B). Figure 13

[0202] In at least one embodiment, AI-assisted annotation 1310 can be used in some examples to generate ground truth data. In at least one embodiment, AI-assisted annotation 1310 (e.g., implemented using an AI-assisted annotation SDK) can utilize a machine learning model (e.g., a neural network) to generate suggested or predicted ground truth data for a customer dataset. In at least one embodiment, a user 1510 can use annotation tools within a user interface (graphical user interface (GUI)) on computing device 1508.

[0203] In at least one embodiment, user 1510 can interact with a GUI via computing device 1508 to edit or fine-tune annotations or automated annotations. In at least one embodiment, a polygon editing feature can be used to move vertices of a polygon to more precise or fine-tuned locations.

[0204] In at least one embodiment, once customer dataset 1506 has associated ground truth data, the ground truth data (e.g., from AI-assisted annotation, manual labeling, etc.) can be used during model training 1314 to generate refined model 1512. In at least one embodiment, customer dataset 1506 can be applied to initial model 1504 any number of times, and ground truth data can be used to update parameters of initial model 1504 until an acceptable level of accuracy is reached for refined model 1512. In at least one embodiment, once refined model 1512 is generated, refined model 1512 can be deployed within one or more deployment pipelines 1410 at a facility for performing one or more processing tasks with respect to medical imaging data.

[0205] ​In at least one embodiment, a refined model 1512 can be uploaded to pre-trained models 1406 in model registry 1324 for selection by another facility. In at least one embodiment, his process can be completed at any number of facilities such that a refined model 1512 can be further refined any number of times on new data sets to generate more general purpose models.

[0206] Figure 15B is an example illustration of a client-server architecture 1532 for augmenting annotation tools with pre-trained annotation models, in accordance with at least one embodiment. In at least one embodiment, an AI-assisted annotation tool 1536 can be instantiated based on client-server architecture 1532. In at least one embodiment, annotation tools 1536 in an imaging application can assist radiologists, for example, in identifying organs and abnormalities. In at least one embodiment, an imaging application can include software tools that help a user 1510 identify a few extreme points on a particular organ of interest in a raw image 1534 (e.g., in a 3D MRI or CT scan), for example, and receive automatic annotation results for all 2D slices of a particular organ. In at least one embodiment, results can be stored as training data 1538 in a data store and used as ground truth data for training, for example, but not by way of limitation. In at least one embodiment, when a computing device 1508 sends extreme points for AI-assisted annotation 1310, a deep learning model, for example, can receive that data as input and return an inference result that segments an organ or abnormality. In at least one embodiment, a pre-instantiated annotation tool (e.g., AI-assisted annotation tool 1536B in Figure 15B AI-assisted annotation tool 1536B) can be augmented by making API calls (e.g., API call 1544) to a server, such as an annotation helper server 1540, which can include a set of pre-trained models 1542 stored in an annotation model registry, for example. In at least one embodiment, an 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 a particular organ or abnormality. In at least one embodiment, these models can be further updated by using a training pipeline 1404. In at least one embodiment, as new labeled clinical data 1312 is added, a pre-installed annotation tool can be improved over time.

[0207] Such components can be used to perform unsupervised domain transfer for one or more neural networks, which can include a feature extraction network and a classifier network.

[0208] 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 above in detail. It should be understood, however, that there is no intention to limit the disclosure to the specific form or forms disclosed, but on the contrary, the intention is to cover all modifications, alternative constructions, and equivalents falling within the spirit and scope of the disclosure, as defined in the appended claims.

[0209] Unless otherwise stated, or as is clear from the context, the use of terms such as "a," "an," and "the" in the context of describing the disclosed embodiments (especially in the context of the following claims) are to be construed to cover both the singular and the plural, unless otherwise indicated by the context. The terms "including," "has," "having," and "comprises" and variations thereof, are to be construed as open-ended terms (meaning "including, but not limited to") unless otherwise noted or as is clear from the context. The term "connected" (when used, without modification, to refer to a physical connection) is to be construed as partly or fully encompassing, attached to, or joined together, even if there are some intervening materials. Unless otherwise indicated herein, the reference within this document to a numerical range is intended merely as a shorthand method of referring individually to each separate value falling within the range, and each separate value is incorporated in the specification as if it were individually recited herein. Unless otherwise indicated or as is clear from the context, the use of the term "set" (e.g., "set of items") or "subset" is to be construed as a non-empty set of one or more members. Furthermore, unless otherwise indicated or as is clear from the context, the term "subset" of a corresponding set does not necessarily denote a proper subset of the corresponding set, but rather the subset and the corresponding set can be equal.

[0210] Unless explicitly stated otherwise or apparent from context, a phrase such as "at least one of A, B, and C" or "at least one of A, B, or C" shall indicate that A is optionally present, with or without B and / or C. In other words, A, B, and C are each optional, independently of one another. For example, "at least one of A, B, and C" or "at least one of A, B, or C" shall mean, in the terms of sets, that A, B, and C can each be present, individually or in any combination, and that there is no need for a combination of A, B, and C to be present if only some of them are present. In an illustrative example, a set with three members: {A}, {B}, and {C}, has the following subsets: {A}, {B}, {C}, {A, B}, {A, C}, {B, C}, {A, B, C}. As such, the phrase "at least one of A, B, and C" or "at least one of A, B, or C" is typically intended to mean any of the possible subsets of A, B, and C, including just A, or just B, or just C, or any combination thereof. Similarly, the phrases "at least one of A, B, and C" or "at least one of A, B, or C" are typically intended to mean any of the possible subsets of A, B, and C, including just A, or just B, or just C, or any combination thereof, unless otherwise indicated or unless it would be clear from context that the phrase is used in an alternative sense. In addition, the phrase "based on" is used to mean "based, at least in part, on" unless otherwise indicated or unless it would be clear from context that the phrase is used in an alternative sense.

[0211] The operations of a process described herein can be performed in any suitable order unless otherwise indicated herein or otherwise clearly contradicted by context. In at least one embodiment, a process, such as those described herein (or variations and / or combinations thereof) is performed under the control of one or more computer systems configured with executable instructions to perform the operations of the process, and the process is implemented as code (e.g., executable instructions, one or more computer programs or one or more applications) that is collectively executed by a hardware or combination of hardware and / or software. In at least one embodiment, the code is stored on a computer-readable storage medium, such as a computer program stored on a computer-readable storage medium, which includes instructions that are executable by one or more processors. In at least one embodiment, a computer-readable storage medium is a non-transitory computer-readable storage medium that excludes transitory signals (e.g., a propagating transient electric or electromagnetic transmission) but includes non-transitory data storage circuitry (e.g., buffers, cache, 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) having stored thereon executable instructions that, as a result of being executed by one or more processors of a computer system (i.e., as a result of being executed), cause the computer system to perform operations described herein. In at least one embodiment, a set of non-transitory computer-readable storage media includes multiple non-transitory computer-readable storage media, and one or more of the individual non-transitory storage media of the multiple non-transitory computer-readable storage media lack all of the code, with the multiple non-transitory computer-readable storage media collectively storing the entire code. In at least one embodiment, executable instructions are executed to cause different instructions to be executed by different processors, e.g., a non-transitory computer-readable storage medium stores instructions and a main central processing unit (“CPU”) executes some instructions and a graphics processing unit (“GPU”) executes other instructions. In at least one embodiment, different components of a computer system have separate processors and different processors execute different subsets of the instructions.

[0212] Accordingly, in at least one embodiment, a computer system is configured to implement one or more services that individually or collectively perform operations of processes described herein, and such a computer system is configured with applicable hardware and / or software to enable implementation of the operations. Moreover, a computer system implementing 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 operating in different manners such that the distributed computer system performs operations described herein and such that a single device does not perform all of the operations.

[0213] The use of any and all examples, or exemplary language (e.g., "such as") provided herein, is intended merely to better illuminate embodiments of the disclosure and does not pose a limitation on the scope of the disclosure unless otherwise claimed. No language in the specification should be construed as indicating any non-claimed element as essential to the practice of the disclosure.

[0214] 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.

[0215] In the description and claims, the terms "coupled" and "connected," along with derivatives thereof, can be used. It should be understood that these terms are not intended as synonyms for each other. Rather, in particular embodiments, "connected" or "coupled" can be used to indicate that two or more elements are in direct or indirect physical or electrical contact with each other. "Coupled" can also mean that two or more elements are not in direct contact with each other, but yet still co-operate or interact with each other.

[0216] Unless specifically stated otherwise, it can be appreciated that throughout the specification terms such as "processing," "computing," "calculating," "determining," or the like, refer to the action and / or processes of a computer or computing system, or similar electronic

[0217] In a similar manner, the term "processor" can refer to any device or portion of a device that processes electronic data from registers and / or memory to transform that electronic data into other electronic data that can be stored in registers and / or memory. As a non-limiting example, a "processor" can be a CPU or GPU. A "computing platform" can include one or more processors. As used herein, a "software" process can include, for example, software and / or hardware entities such as tasks, threads, and intelligent agents that perform work over time. Likewise, each process can refer to multiple processes to sequentially or concurrently execute instructions, either continuously or intermittently. The terms "system" and "method" can be used interchangeably herein so long as a system can embody one or more methods and a method can be considered a system.

[0218] In this document, obtaining, accessing, receiving, or inputting analog or digital data into a subsystem, computer system, or computer-implemented machine can be referenced. Analog and digital data can be obtained, accessed, received, or inputted 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, accessing, 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, accessing, receiving, or inputting analog or digital data can be accomplished by transmitting data from a providing entity to an accessing entity via a computer network. Providing, outputting, transmitting, sending, or presenting analog or digital data can also be referenced. 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, a parameter of an application programming interface, or an interprocess communication mechanism.

[0219] Although the above discussion discloses example implementations of the described technology, other architectures can be utilized and are intended to fall within the scope of the present disclosure. Moreover, although a specific division of responsibilities has been defined above for purposes of discussion, various functions and responsibilities can be distributed and divided in different ways depending on circumstances.

[0220] Further, although the subject matter has been described in language specific to structural features and / or methodological acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described. Rather, the specific features and acts are disclosed as example forms of implementing the claims.

Claims

1. A computer-implemented method, comprising: The first neural network is used to encode a set of features representing the input into the latent space; The input is classified using a second neural network, at least in part, based on encoded features; At least in part, the domain corresponding to the input is inferred based on the encoded features and using a third neural network. as well as At least one or more network parameters of the first neural network are adjusted until the second neural network accurately classifies the input based on features encoded by the first neural network, wherein adjusting the one or more network parameters prevents the third neural network from reasoning about the domain with at least minimal determinism.

2. The computer-implemented method of claim 1, wherein the input comprises one or more images, the one or more images comprising representations of one or more types of objects, and wherein the domain is further inferred to predict objects corresponding to the one or more types of objects.

3. The computer-implemented method of claim 1, wherein the input comprises one or more labeled images from a first domain and one or more unlabeled images from a second domain.

4. The computer-implemented method of claim 1, wherein the network parameters of the first neural network are adjusted such that the first neural network encodes features independent of the domain into the latent space.

5. The computer-implemented method as described in claim 1, further comprising: The network parameters are provided for the first neural network and the second neural network to perform classification inference on unclassified data.

6. The computer-implemented method of claim 1, wherein the first neural network is a convolutional neural network, the convolutional neural network being trained to extract features from input data and encode those features into a latent space.

7. The computer-implemented method of claim 6, wherein the second neural network is a classifier network for taking the latent space as input and inferring the classification of one or more objects or instances represented in the input data.

8. A training system, comprising: Feature extraction networks are used to encode a set of features extracted from the input into a latent space; A data classifier network for classifying the input based at least in part on encoded features; A domain classifier network for inferring the domain corresponding to the input based at least in part on the encoded features; and An evaluation component is used to adjust one or more network parameters of at least one of the feature extraction network, the data classifier network, or the domain classifier network until the data classifier network accurately classifies the input using the features encoded by the feature extraction network, but prevents the domain classifier network from inferring the domain with at least minimal determinism.

9. The training system of claim 8, wherein the network parameters of the feature extraction network are adjusted such that the feature extraction network encodes features independent of the domain into the latent space.

10. The training system of claim 8, wherein the training system further provides the network parameters for the feature extraction network and the data classifier network to infer one or more classifications of unclassified data.

11. The training system of claim 8, wherein the feature extraction network is a convolutional neural network.

12. The training system of claim 8, wherein the data classifier network is a two-layer neural network with a Leaky-Relu activation function.

13. The training system of claim 8, wherein the objective of the f-adversarial learning process implemented by the training system utilizes Stackelberg equilibrium.

14. The training system of claim 8, wherein the input comprises one or more labeled images of one or more objects from a first domain and one or more unlabeled images of one or more objects from a second domain, and wherein the domain classifier network further infers the domain based on at least one predicted object among the one or more objects.

15. A computer-implemented method, comprising: The feature extraction network, the data classifier network, and the domain classifier network are trained using the training system described in claim 8. The trained feature extraction network receives inputs corresponding to domains in a set of domains; The trained feature extraction network extracts a set of domain-independent features from the input. and The trained data classifier network infers the classification of the input based at least in part on the extracted feature set.

16. The computer-implemented method as described in claim 15, wherein, The feature extraction network and the data classifier network are trained using data from the set of domains, the data including unlabeled data from the first domain and labeled data from the second domain.

17. The computer-implemented method of claim 16, wherein the feature extraction network and the data classifier network are trained together with the domain classifier network.

18. The computer-implemented method of claim 17, wherein, The goal of the f-adversarial learning process implemented in the training is to utilize Stackelberg equilibrium.

19. The computer-implemented method of claim 16, wherein the feature extraction network is a convolutional neural network.

20. The computer-implemented method of claim 16, wherein the data classifier network is a two-layer neural network with a Leaky-Relu activation function.

Images