Patch Deployed Deep Neural Networks for Autonomous Machine Applications

Abstract

To provide rapid resolution of deep neural network (DNN) failure modes by deploying patch neural networks (PNNs) trained to operate effectively on the failure modes of the DNN.SOLUTION: PNNs operate on the same or additional data as a DNN, and generate new signals in addition to those generated using the DNN that address failure modes of the DNN. A fusion mechanism is employed to determine which output to rely on for a given instance of the DNN / PNN combination. As a result, failure modes of the DNN are addressed in a timely manner that requires minimal deactivation or downtime for the DNN, a feature controlled using the DNN, and / or semi-autonomous or autonomous functionality as a whole.SELECTED DRAWING: Figure 1A

Description

[Background technology]

[0001] Autonomous and semi-autonomous vehicles utilize deep neural networks (DNNs) to help perform various tasks during operation, such as, without limitation, object detection, classification, and / or tracking, path planning, control decisions, obstacle avoidance, and / or other driving-related tasks. Due to the safety-critical nature of autonomous and semi-autonomous driving operations, these DNNs must be accurate and precise, but also undergo extensive validation and testing prior to deployment in operational vehicles. For example, training, testing, and validation of DNNs can take anywhere from six months to a year or more in some cases, particularly where regulatory and / or original equipment manufacturer (OEM) requirements must be met. For example, OEMs may require that a retrained DNN not result in any change in the system's observed behavior, for example, over thousands of miles during testing of the DNN. This lengthy schedule is in direct opposition to the turnaround time expected by consumers due to safety considerations surrounding the autonomous functionality of a vehicle. For example, consumers may expect failure modes to be addressed within hours or days in order to feel comfortable operating a vehicle using semi-autonomous or autonomous features.

[0002] Conventional approaches address this issue by retraining the DNN using additional training data that includes training data instances that correspond to the failure modes of the DNN. During this retraining and validation period - e.g., six months to a year or more - the DNN may be operated with the failure modes still present or may result in the deactivation of the DNN, its features being controlled using the DNN, and / or semi-autonomous or autonomous functionality as a whole. However, even when these retrained DNNs are used, the retrained DNNs may suffer from new or additional failure modes that were not identified during training, testing, or validation. For example, during retraining to address identified failure modes, new failure modes may surface - e.g., due to unexpected object types or road profiles or pose types, the effects of certain weather conditions, etc. - that did not materialize during testing, training, and validation. Each new failure mode then creates a cascading effect that may result in one or more features of the vehicle being unavailable for an extended period of time - thereby causing a negative user experience and potentially a lack of confidence in the product by consumers - and requiring the DNN to be retrained. As such, these re-validation techniques are lengthy, costly, and generally infeasible for patching failure modes of DNNs deployed in operational environments. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] U.S. Patent Application Serial No. 16 / 101,232 Summary of the Invention [Means for solving the problem]

[0004] An embodiment of the present disclosure relates to training and deployment of patches for deep neural networks (DNNs). Systems and methods are disclosed that enable rapid resolution of DNN failure modes discovered during the deployed operation of the DNN. For example, in addition to or instead of retraining, retesting, and revalidating the DNN each time the failure mode is realized, one or more patch neural networks (PNNs) trained to operate effectively with the DNN in the failure mode may be deployed to address the failure mode - at least temporarily - without requiring the DNN features to be controlled using the DNN and / or semi-autonomous or autonomous functionality as a whole to be deactivated. As a result, downtime of the DNN or other parts of the system may be dramatically reduced compared to conventional systems. Additionally, during deployment of the PNN, an updated version of the DNN may be trained, tested, and validated prior to deployment in the vehicle, at which point the particular PNN used to address the DNN failure mode may no longer be needed.

[0005] The PNN may operate on the same or additional data as the DNN and may generate new signals in addition to those generated using the DNN. For example, in an embodiment, the PNN may be trained to generate outputs that address the failure modes of the DNN - in addition to computing similar or the same output types as the DNN. A fusion mechanism may be used to evaluate the outputs of the DNN and the PNN and / or to evaluate additional criteria to determine which output to rely on for a given instance of a DNN / PNN combination. During training, the parameters of the PNN (e.g., weights and biases) may be trained using a combination of the initial training data used to train the DNN plus new training data selected to address the failure modes of the DNN. The parameters of the DNN may be fixed or frozen during the training of the DNN so that the initial performance of the DNN is maintained while allowing the PNN to recreate the outputs of the DNN in addition to the new or updated outputs of the PNN. The introduction of a PNN may require additional computation due to the additional nodes and / or layers of the PNN required to process the input data. As such, the PNN may be subject to pruning during training to remove entire layers and / or nodes of a layer to reduce the computational and latency impact of the PNN on the system.

[0006] In some embodiments, after being trained, the PNN and / or fusion mechanism: The code may be pushed over the air (OTA) to consumer vehicles and run in a passive or inactive mode, where the code is executed but not activated or acted upon. Data recorded during this (e.g., passive) mode may be evaluated to verify the performance of the PNN and / or fusion mechanisms or to indicate where additional training is needed. As such, training and validation of PNN functionality in the system may be expedited by leveraging a consumer fleet of vehicles already running deployed DNNs - thereby increasing the likelihood of meeting rigorous testing and validation standards for PNNs as well as fast turnaround time expectations of consumers.

[0007] The present systems and methods for training and deploying patches for deep neural networks (DNNs) are described in detail below with reference to the accompanying drawings. [Brief description of the drawings]

[0008] [Figure 1A] 1 is a data flow diagram illustrating an example process for executing a patch neural network (PNN) during deployment, according to some embodiments of the present disclosure. [Figure 1B] 1B illustrates exemplary deep neural network (DNN) and PNN architectures for processing input data within the process of FIG. 1A, in accordance with some embodiments of the present disclosure. [Figure 1C] FIG. 1 illustrates an example layer implementation of a DNN, a PNN, or a combination thereof, according to some embodiments of the present disclosure. [Figure 2A] FIG. 1 illustrates an example architecture of a combined PNN and DNN, in accordance with some embodiments of the present disclosure. [Figure 2B] FIG. 1 illustrates an example architecture of a combined PNN and DNN, in accordance with some embodiments of the present disclosure. [Figure 3A]FIG. 1 illustrates an example DNN and PNN architecture with selective activation of the PNN, in accordance with some embodiments of the present disclosure. [Figure 3B] FIG. 1 illustrates an example DNN and PNN architecture with selective activation of the PNN, in accordance with some embodiments of the present disclosure. [Figure 4] 1 is a flow diagram illustrating a method for executing a PNN during deployment, according to some embodiments of the present disclosure. [Figure 5A] 1 is an illustration of an example autonomous vehicle, according to some embodiments of the present disclosure. [Figure 5B] 5B is an illustration of camera positions and fields of view for the example autonomous vehicle of FIG. 5A, according to some embodiments of the present disclosure. [Figure 5C] FIG. 5B is a block diagram of an example system architecture of the example autonomous vehicle of FIG. 5A, in accordance with some embodiments of the present disclosure. [Figure 5D] FIG. 5B is a system diagram of communication between a cloud-based server and the example autonomous vehicle of FIG. 5A, according to some embodiments of the present disclosure. [Figure 6] FIG. 1 is a block diagram of an example computing device suitable for use in implementing some embodiments of the present disclosure. [Figure 7] FIG. 1 is a block diagram of an exemplary data center suitable for use in implementing some embodiments of the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] Systems and methods are disclosed for training and deployment of patches for deep neural networks (DNNs). The present disclosure may be described with respect to an exemplary autonomous vehicle 500 (alternatively referred to herein as "vehicle 500" or "ego vehicle 500", examples of which are described with reference to FIGS. 5A-5D), but this is not intended to be limiting. For example, the systems and methods described herein may be used by, without limitation, non-autonomous vehicles, semi-autonomous vehicles (e.g., in one or more adaptive driver assistance systems (ADAS)), piloted or non-piloted robots or robotic platforms, warehouse vehicles, off-road vehicles, vehicles coupled to one or more trailers, airships, boats, shuttles, emergency response vehicles, motorcycles, electric or mopeds, aircraft, construction vehicles, submarines, drones, and / or other vehicle types. Additionally, while the present disclosure may be described with respect to semi-autonomous or autonomous machine functionality, this is not intended to be limiting, and the systems and methods described herein may be used in augmented reality, virtual reality, mixed reality, robotics, security and surveillance, autonomous or semi-autonomous machine applications, medical imaging applications, language processing applications, smart home applications, deep learning applications, and / or any other technology space where neural networks or other types of machine learning models may be used.

[0010] Referring to FIG. 1A, FIG. 1A is a data flow diagram illustrating an exemplary process for executing a patch neural network (PNN) 104B during deployment, according to some embodiments of the present disclosure. It should be understood that this and other arrangements described herein are described merely as examples. Other configurations and elements (e.g., machines, interfaces, functions, sequences, groupings of functions, etc.) may be used in addition to or instead of those illustrated, and some elements may be omitted entirely. Furthermore, many of the elements described herein are functional entities that may be implemented as separate or distributed components or in conjunction with other components, and in any suitable combination and location. Various functions described herein as being performed by an entity may be implemented by hardware, firmware, and / or software. For example, various functions may be implemented by a processor executing instructions stored in a memory. In some embodiments, training and / or deployment of the deep neural network (DNN) 104 (e.g., including the PNN 104B and / or the base DNN 104A) may be performed, at least in part, using similar components, features, and / or functionality as described herein with respect to the vehicle 500 of Figures 5A-5D, the example computing device 600 of Figure 6, and / or the example data center 700 of Figure 7. The DNN 104 may be referred to herein as a combined DNN 104, e.g., including a combination of the base DNN 104A and the PNN 104B.

[0011] The process 100 may include generating and / or receiving input data 102 from one or more sources (e.g., sensors of the vehicle 500, a robot, a VR, AR, or mixed reality system, a medical imaging device, a microphone of a device for audio processing, and / or another source type). In some examples, the input data 102 (alternatively referred to as “sensor data 102”) may be received from one or more sensors of a vehicle (e.g., the vehicle 500 of FIGS. 5A-5D described herein), by way of non-limiting example. The sensor data 102 may be used by the vehicle 500 in the process 100 during training and / or deployment of the one or more base DNNs 104A and / or the one or more PNNs 104B to perform any number of operations, such as object detection, classification, and / or tracking, path planning, control operations, intersection structure detection, road layout determination, mapping and localization, etc. During training, the sensor data 102 may be generated using one or more data collection vehicles that generate sensor data for training the DNN, e.g., DNN 104, and / or may be pre-generated and included in a training data set. The sensor data 102 used during training may additionally or alternatively be generated using simulated sensor data (e.g., sensor data generated using one or more virtual sensors of a virtual vehicle in a virtual environment) and / or augmented sensor data (e.g., sensor data generated using one or more data collection vehicles and modified with virtual data, respectively). Once trained and deployed in the vehicle 500, the sensor data 102 may be generated by one or more sensors of the vehicle 500 and processed by the DNN 104 to compute various outputs 106, as described herein.

[0012] In some embodiments, such as during training of the PNN 104B, the training sensor data may include further additional and / or alternative training data used to train the base DNN 104A to account for one or more failure modes of the base DNN 104A that the PNN 104B was trained to address. For example, a failure mode of the base DNN 104A in an end user's product or system may be identified and the input data 102 that resulted in the failure mode may be collected and used to train the PNN 104B. In some embodiments, the shadow modes of the base DNN 104A and / or the PNN 104B may be used to identify input data instances when the base DNN 104A experienced a failure and / or when the output 106 of the PNN 104B is determined to be more reliable or accurate. These identified input data instances corresponding to failure modes of the base DNN 104A may also be used to derive additional training input data collections from one or more data collection vehicles, from a consumer fleet of vehicles, and / or within a simulation environment (e.g., if the base DNN 104A has a failure mode related to an electric scooter, a simulation may be performed including an electric scooter to generate synthetic training data instances for training the PNN 104B).

[0013] 5A-5C, the sensor data 102 may include, but is not limited to, sensor data 102 from any of the sensors of the vehicle 500, including, for example, and with reference to Figures 5A-5C, a RADAR sensor 560, an ultrasonic sensor 562, a LIDAR sensor 564, a stereo camera 568, a wide-view camera 570 (e.g., a fish-eye camera), an infrared camera 572, a surround camera 574 (e.g., a 360-degree camera), a long-range and / or mid-range camera 578, a microphone 596, an IMU sensor 566, a GNSS sensor 558, a speed sensor 544, and / or other sensor types. As another example, the sensor data 102 may include virtual (e.g., simulated or augmented) sensor data generated from any number of sensors of a virtual vehicle or other virtual object in a virtual (e.g., test) environment. In one such instance, the virtual sensors may correspond to virtual vehicles or other virtual objects in a simulated environment (e.g., used to test, train, and / or validate DNN performance) and the virtual sensor data may represent sensor data captured by the virtual sensors in the simulated or virtual environment. As such, by using virtual sensor data, the DNNs 104 described herein may be tested, trained, and / or validated using simulated or augmented data in a simulated environment, which may enable testing of more extreme scenarios outside of real-world environments where such testing may be less safe.

[0014] In some examples, the sensor data 102 may include image data representing an image, image data representing a video (e.g., a snapshot of a video), and / or sensor data representing a representation of a sensor's field of perception (e.g., a depth map for a LIDAR sensor, a value graph for an ultrasonic sensor, etc.). When the sensor data 102 includes image data, any type of image data format may be used, including, but not limited to, compressed images, such as Joint Photographic Experts Group (JPEG) or Luminance / Chrominance (YUV) formats, compressed images, such as frames resulting from compressed video formats, such as H.264 / Advanced Video Coding (AVC) or H.265 / High Efficiency Video Coding (HEVC), raw images, such as those derived from Red Clear Blue (RCCB), Red Clear (RCCC), or other types of image sensors, and / or other formats. Additionally, in some instances, the sensor data 102 may be used in the process 100 without preprocessing (e.g., in a raw or captured format), while in other instances, the sensor data 102 may undergo preprocessing (e.g., noise balancing, demosaicing, scaling, cropping, enhancement, white balancing, tone curve adjustment, etc., e.g., using a sensor data preprocessor (not shown)). As used herein, sensor data 102 may refer to raw sensor data, preprocessed sensor data, or a combination thereof.

[0015] The input data 102 used to train the base DNN 104A and / or the PNN 104B may include initial input data (e.g., as captured by one or more image sensors), downsampled input data, upsampled input data, cropped or region of interest (ROI) input data, otherwise augmented input data, and / or combinations thereof. The DNN 104 may be trained using one or more loss functions and using the input data 102 as well as corresponding ground truth data until the DNN 104 converges to an acceptable or optimal level of accuracy.

[0016] In embodiments in which the DNN 104 - e.g., the base DNN 104A and / or the PNN 104B - is implemented outside of the vehicle 500, the input data 102 may include any type of data that the DNN may be capable of processing. For example, the input data 102 may include audio data (e.g., spectrograms), image data, text data, location data, sensor data (e.g., from LiDAR sensors, RADAR sensors, etc.), and / or other data types that can be represented in a suitable format (e.g., as an input vector or tensor) for processing by the DNN.

[0017] The input data 102 may be applied to one or more DNNs 104, which may include the base DNN 104A and the PNN 104B. For example, the base DNN 104A may include an initial or previously trained DNN - after deployment or during operation - that has one or more failure modes. For example, with respect to object detection, a failure mode may occur when the base DNN 104A is trained to identify vulnerable road users (VRUs). In one such example, after being deployed, the base DNN 104A may be determined to be not accurate enough when faced with a VRU that includes a person riding a scooter - e.g., an electric scooter. As such, the PNN 104B may be trained as a dedicated electric scooter detector. As another example, for example, with respect to patching a particular section of a road that is difficult for the vehicle 500 to cross based on the output of the base DNN 104A, the PNN 104B may be trained to address the particular section of the road and / or multiple road sections similar to the particular section of the road. As such, when an existing signal 112 may be received, for example, from a mapping or navigation system indicating that a vehicle 600 is on or approaching that section of road, the output of the PNN 104B may be used to account for inaccuracies or failure modes of the base DNN 104A. As a further example, with respect to object classification, the base DNN 104A trained as a vehicle classifier may have difficulty accurately classifying a new car model, such as a future emerging truck. In this example, the PNN 104B may be trained to accurately classify the new car model.

[0018] Although examples are described herein with respect to the use of neural networks as the DNN 104, this is not intended to be limiting. For example, and without limitation, the DNN 104 described herein may include any type of machine learning model, such as linear regression, logistic regression, decision trees, support vector machines (SVMs), naive Bayes, k-nearest neighbors (Knn), K-means clustering, random forests, dimensionality reduction algorithms, gradient boosting algorithms, machine learning models using neural networks (e.g., autoencoders, convolutional, recurrent, perceptrons, long / short term memory (LSTM), Hopfield, Boltzmann, deep belief, deconvolution, generative adversarial, liquid state machines, etc.), and / or other types of machine learning models. The PNN 104B may include the same type of DNN as the base DNN 104A, or may include a different type of DNN.

[0019] The base DNN 104A may be trained, tested, and validated on a training data set, and after validation, may be deployed for use in the vehicle 500. Once a failure mode has been identified, the PNN 104B may be trained to remedy or account for the failure mode of the base DNN 104A, in an embodiment, by undergoing training using some or all of the training data from the training data set used to train the base DNN 104A along with additional or alternative training data generated or obtained to account for the failure mode of the base DNN 104A. During training of the PNN 104B, the parameters (e.g., weights and biases) of the base DNN 104A may be fixed such that the base DNN 104A is not changed from its previous validation state. In addition, the relationship between the output of the nodes of the PNN 104B and the input of the nodes of the base DNN 104A may be fixed or set to 0 such that the base DNN 104A does not process the output of the layers or nodes of the PNN 104B. In contrast, parameters related to the relationship between the outputs of the nodes of the base DNN 104A and the inputs of the nodes of the PNN 104B can be trained such that the PNN 104B processes the outputs of the nodes or layers of the base DNN 104A in addition to the nodes or layers of the PNN 104B.

[0020] For example, as shown in FIG. 1B, the input data 102 may be applied to the layers 126 (e.g., input layers 126A-1 and 126B-1) of both the base DNN 104A and the PNN 104B, and the output of the layer 126A of the base DNN 104A may be passed to the input of the next layer 126A and the input of the next layer 126B of the PNN 104B, while the output of the layer 126B of the PNN 104B may be passed only to the input of the next layer 126B of the PNN 104B. An expanded view of the DNN / PNN combination of FIG. 1B is shown in FIG. 2A. In such an example, the integrity of the base DNN 104A may be preserved, while the PNN 104B may only address the failure modes of the base DNN 104A. The PNN 104B can predict, in an embodiment, each of the outputs 106A of the base DNN 104A and / or additional or alternative outputs 106B beyond those of the base DNN 104A. For example, if the base DNN 104A is trained to calculate confidence values ​​for a vehicle class and a bicyclist class, the PNN 104B can be trained to calculate confidence values ​​for the vehicle class, the bicyclist class, and an additional class, such as an electric scooter class, of the PNN 104B. As such, if the fusion mechanism 108 indicates that the outputs 106B of the PNN 104B should be used for a given iteration or time step, the outputs 106B can include those of the base DNN 104A and those of the PNN 104B. In one such example, an additional, modified, or alternative output of PNN 104B should be relied upon, but in instances where the fusion mechanism 108 indicated an incorrect output to rely on for a given iteration or time step by including an instance of the output of base DNN 104A in the output 106B of PNN 104B, the fused output 114 may still be accurate (e.g., if the fusion mechanism 108 properly selected the base DNN 104A for a given iteration, the fused output 114 may still reflect the output 106A of the base DNN 104A).As such, until the updated base DNN 104A has undergone a full re-validation process, the PNN 104B may be trained to run a superior - failure mode corrected - version of the base DNN 104A that may be implemented only for failure modes of the base DNN 104A due to regulatory and / or OEM testing and validation requirements.

[0021] However, in some embodiments, the PNN 104B may be trained to calculate a unique output 106B with respect to the output 106A of the base DNN 104A. Furthermore, in embodiments such as those described herein with respect to Figures 2B and 3A, a single output may be calculated using the base DNN 104A (e.g., when the PNN 104B is selectively turned off or otherwise deactivated based on the determination of the fusion mechanism 108) or a combination of the base DNN 104A and the PNN 104B (e.g., when the PNN 104B is selectively turned on or otherwise activated based on the determination of the fusion mechanism 108).

[0022] Although the diagrams of Figures 1B, 2A-2B, and 3A-3B include layers 126B of the PNN 104B corresponding to each layer 126A of the base DNN 104A, this is not intended to be limiting. For example, each layer 126A of the base DNN 104A may not have a corresponding layer 126B in the PNN 104B. As such, one or more initial layers 126A of the base DNN 104A may not include a corresponding layer 126B of the PNN 104B, and the PNN layer 126B may start at any point along the sequence of layers 126A of the base DNN 104A. As a non-limiting example, it may be determined that a feature extractor layer 126A of the base DNN 104A - e.g., a layer 126A of the base DNN 104A that calculates a feature map - does not require that the corresponding layer 126B of the PNN 104B is used exactly in the calculation of the output 106B of the PNN 104B. As such, layer 126B of PNN 104B may not be included until after feature extractor layer 126A of base DNN 104A. After the PNN layer is introduced, each next layer of the combined base DNN / PNN network may include both the PNN layer 126B and the base DNN layer 126A.

[0023] Referring to FIG. 1C, FIG. 1C illustrates implementations of various layers of the base DNN 104A and the PNN 104B. Each diagram in FIG. 1C corresponds to a single layer, which may be an input layer, an output layer, or a layer between an input layer and an output layer. For example, the base DNN implementation 120 illustrates one implementation of a single layer 126A of the base DNN 104A, e.g., having weights, W,, that receives an input (e.g., vector or tensor), X, 124A (e.g., an input corresponding to the input data 102 to the input layer 126A of the base DNN 104A or an input from a previous layer 126A of the base DNN 104A) and calculates an output (e.g., vector or tensor), Y, 128A. The weights, W, (and / or other parameters) of this layer 126A may be pre-trained prior to the deployment of the base DNN 104A in the vehicle 500.

[0024] The DNN implementation 122A (e.g., a combined base DNN / PNN implementation) shows an implementation of a single layer 126A of the base DNN 104A, e.g., with weights, W, that receives an input, X, 124A, and calculates an output, Y, 128A, and a single layer 126B of the PNN 104B, e.g., with weights, W', that receives an input, X, 124A and an input, X', 124B, and calculates an output, Y', 128B. As such, two separate vector or tensor by matrix multiplication operations may be performed. In some instances, such as when the layer 126B of the PNN 104B corresponds to the first or input layer of the base DNN 104A, the X' input may include additional input data 102 other than that provided as an input to the base DNN 104A. For example, the input data 102 of the base DNN 104A may include an image, and the input data 102 of the PNN 104B may include an image and / or another input type, such as another image, a signal from a different sensor (e.g., a speed sensor, a position sensor, etc.), and / or another input data type. In other instances, such as when the layer 126B of the PNN 104B corresponds to a layer other than the input layer or the first layer of the base DNN 104A, the X' input may correspond to the Y' output of the previous layer 126B of the PNN 104B. For example, after the first PNN layer 126B is introduced, the first PNN layer 126B may output a different output, Y', than the base DNN layer 126A, and thus the next PNN layer 126B may receive the Y' output as the X' input. As such, the PNN layer 126B may receive the input, X, that the base DNN layer 126A receives in addition to the input, X'.

[0025] A DNN implementation 122A (e.g., a combined base DNN / PNN implementation) includes adding a corresponding additional layer 126B to an existing layer 126A. A PNN 104B using separate layers similar to those of the implementation 122A may be relatively small and may receive an input concatenation of X and X' and an output Y'. During the training process, only the patch layer 126B may be trained.

[0026] The DNN implementation 122B shows an example implementation of a single combined layer 126C that includes an extension layer that accounts for both the layer 126A of the base DNN 104A and the layer 126B of the PNN 104B, with weights W, W', and zeros 132 corresponding to the relationship between the output of the node or the value of the PNN 104B and the input to the node or the value of the base DNN 104A. As such, the combined layer 126 can receive concatenated inputs (e.g., vectors or tensors) 124A and 124B representing X and X', and can output concatenated outputs (e.g., vectors or tensors) 128A and 128B representing Y and Y'. The zeros 132 can increase the dimensionality of the initial layer 126A of the base DNN 104A to match that of the concatenated inputs including X and X', and can represent the weights of the relationship between the additional input X' and the initial output Y. As such, a single vector or tensor by matrix multiplication operation may be performed, but the results may be the same as using the DNN implementation 122A.

[0027] During the training process, only a subset of the weights, W', corresponding to PNN 104B may be trained. To keep the initial output, Y, unchanged, all relationships between the additional input, X', and the initial output, Y, are removed (e.g., the weights of those relationships are set and maintained at zero).

[0028] The implementations 122A and 122B may be equivalent in terms of the resulting DNN functionality. In addition, W' may be the same in both implementations, so that simple transition between the two versions may be realized. For example, the first implementation 122A may be used to train the weights W', and the second implementation 122B may be used in the deployment of the DNN 104. The advantage of the implementation 122A is that the X inputs of all layers may be precomputed using the initial base DNN 104A, and only the PNN 104B may be considered during training - for example, since layer 126B is separate from layer 126A of the base DNN 104A. This may increase the training process of the PNN 104B, but may require additional storage and bandwidth. In addition, because the layers are separate, layer optimization may be less effective, thereby potentially resulting in slower inference than when the implementation 122B is used. Advantages of the second implementation 122B include compatibility with optimization tools (e.g., NVIDIA's TensorRT optimization tool). For example, new or additional layers are not added - e.g., only existing layers are extended - so there are fewer optimization problems with model inference. Disadvantages of implementation 122B are the requirement to store the zeroed weights (e.g., minimal increase to model size) and the need for the entire DNN, including the combined PNN nodes and base DNN nodes, to be used in training PNN 104B (which may be overcome by training weights W' using implementation 122A and converting them to implementation 122B).

[0029] In any example, after the PNN 104B is trained, one or more pruning operations may be performed to remove layers or nodes of the PNN 104B that are unnecessary in the computation of the accurate output 106B. For example, a low-level feature extractor layer may not need to be patched, and thus a PNN layer 126B corresponding to a low-level feature extractor layer 126A of the base DNN 104A may be removed - or its nodes may be removed - to reduce the size of the layer 126B. As such, during training, outputs before and after various pruning operations may be compared to determine a tradeoff between size and accuracy of the PNN 104B.

[0030] In some instances, two or more PNNs 104B may be trained to patch the base DNN 104A. For example, an additional failure mode may surface after the first PNN 104B is deployed, and another or alternative PNN 104B may be trained for the additional and / or previous failure mode. In some embodiments, the second PNN 104B may include a patch of the first PNN 104B in addition to a patch of the new failure mode - to avoid stacking of the PNNs 104B. For example, the second PNN 104B may be trained to patch the failure mode of the existing PNN 104B deployed in the vehicle 500 in addition to another failure mode, so that the existing PNN 104B may be removed and a new or updated PNN 104B may replace the existing PNN 104B to account for both failure modes. However, in an embodiment, two or more PNNs 104B may be deployed to patch a single base DNN 104A. In some embodiments, after the new or updated PNN 104B is deployed, the fusion mechanism 108 may also require updating. For example, a new or updated gating NN 110 may be trained to account for the new or updated PNN 104B.

[0031] 1A, the fusion mechanism 108 can use the outputs 106--e.g., the output 106A from the base DNN 104A and the output 106B from the PNN 104B--to determine the fused output 114. For example, the fusion mechanism 108 can use values ​​from the outputs 106A and 106B (e.g., confidence values ​​corresponding to object types when the DNN 104 is an object classification DNN), outputs from one or more gating NNs 110, and / or existing signals 112 (e.g., corresponding to geographic, weather, lighting, or other information) to determine which of the outputs 106A or 106B to rely on for a given iteration. The fusion mechanism 108 can be a rule-based (e.g., non-learned) decision mechanism, a learned decision mechanism, or a combination thereof. The selection of the fusion mechanism (e.g., the output of the base DNN 104A and the PNN 104B, the existing signal 112, or the confidence from the gating NN 110) may be based on the failure mode type of the base DNN 104A for which the PNN 104B was trained. If the gating NN 110 is used, the gating NN 110 may be trained to detect the failure mode or scenario. The gating NN 110 may have access to the input data 102 of the base DNN 104A and / or the PNN 104B as well as, in some embodiments, the internal state (not shown) of the DNN 104. For example, the gating NN 110 may be trained as a classifier with data from both new failure mode scenarios as well as the initial data to distinguish between the two data types. Due to the simplicity of the gating NN 110 - e.g., trained only to detect a specific data type - the gating NN 110 may be implemented as a small network that is fast to train with high performance on only a small amount of training data.

[0032] For example, if the PNN 104B is trained for a location based on a failure (e.g., inability to handle a particular intersection, exit ramp, roadway, etc.), the existing signal 112 may be used. For example, GNSS data may be used to indicate when a vehicle 500 is approaching or located at a location that the PNN 104B was trained to patch. As another example, a vehicle detector may fail to detect a new model or prototype vehicle due to its unique or unusual shape. In this type of example, the fusion mechanism 108 may use confidence from the outputs of the base DNN 104A and the PNN 104B and / or may use a gating NN 110 that analyzes the input data 102 (e.g., the same input data 102 as the base DNN 104A and / or the PNN 104B in an embodiment) and outputs a binary output that indicates which output to use. For example, if the confidence of the PNN 104B is high for the vehicle type and the confidence of the base DNN 104A is low for each of its classes, the output of the PNN 104B may be relied upon. As another example, the gating NN 110 may be trained to output a value of “1” when a future vehicle is detected, such that the output of the PNN 104B may be relied upon as the fused output 114 for a given iteration.

[0033] As such, different fault types may require different patching strategies, including the type of PNN 104B and the fusion mechanism 108 used therein. A few non-limiting example use cases of the PNN 104B for autonomous or semi-autonomous driving may include adding new features, patching specific sections of roads, or improving existing features. For example, to add new features, the base DNN 104A may be deployed before e-scooters are widespread. E-scooter riders may be considered vulnerable road users (VRUs), and the base DNN 104A, trained to identify or detect VRUs, may be determined to have failed to detect e-scooter riders as VRUs. In this example, the PNN 104B may be trained as a dedicated e-scooter detector, and the fusion mechanism 108 may be coupled with other VRU detectors of the base DNN 104A, such as pedestrians, bicyclists, etc.

[0034] With respect to patching a particular section of road, the base DNN 104A deployed to aid in lane keeping or route planning through roads and / or intersections may have a failure mode at a particular section or road or intersection - e.g., due to insufficient road markings, unique intersection design, etc. The vehicle 500 may not be able to properly keep lane while traversing the driving surface. In this instance, the PNN 104B may be trained and validated for lane keeping on the particular section of road that the base DNN 104A addresses. In one such instance, the fusion mechanism 108 may include GNSS signals, localization to a high definition (HD) map (e.g., using sensor data from sensors of the vehicle 500), and / or training of a dedicated landmark detector for the particular section of road or intersection. As such, the output 106B of the PNN 104B may be used when a signal is received indicating that the vehicle 500 is on the section of road corresponding to the PNN 104B.

[0035] If an existing feature, e.g., a vehicle detector, is to be improved, the base DNN 104A may become inaccurate in predicting a particular type of vehicle - e.g., a new model or design. In one such example, training a dedicated PNN 104B for new passenger vehicle detection may not work, since the appearance of the new vehicle model may be too close to similar models, and it may be difficult to force the PNN 104B to distinguish. Also, simply concatenating the outputs as in the example of adding new features may not be feasible, since validation of the false positive rate of the PNN 104B may not be possible without a full revalidation of the base DNN 104A. As a result, both the output 106A of the base DNN 104A and the output 106B of the PNN 104B may be used to make a decision, since neither the false positive rate nor the confidence output of the PNN 104B alone may be sufficient to reach the required safety level. However, the combination of the two outputs 106 may achieve an adequate level of safety - for example, a false positive in the final output signal may occur only if the fusion mechanism 108 fails and the PNN 104B has a false positive.

[0036] Referring to FIG. 2B, an alternative architecture to the base DNN / PNN combination of FIG. 1B and 2A is shown. In one such example, the nodes of PNN 104B may be added to the inputs of existing nodes of base DNN 104A. Thus, PNN 104B may not need to generate its own output, since it may learn to correct existing DNN outputs of failure mode. When PNN 104B is turned off, base DNN 104A may run exactly as it did before patching. However, since both base DNN output 106A and PNN output 106B are not calculated at each iteration, if both outputs are needed, inference may be required to run twice to obtain the two outputs of the analysis, which increases computation and increases latency.

[0037] 3A-3B, which show examples of DNN and PNN architectures with selective activation of PNNs. For example, with respect to FIG. 3A, the fusion mechanism 108 may determine whether to turn on or activate a node of the PNN 104B based on the existing signal 112 and / or the gating NN 110. For example, when a node of the PNN 104B is turned off or deactivated, the values ​​of the parameters (e.g., weights and biases) of each of the nodes of the PNN 104B may be fixed to zero. This ensures that the bit-accurate-based DNN 104A operates exactly as it did before the PNN 104B was added. Turning on or activating the nodes of the DNN 104B allows the trained parameters of the PNN 104B to operate normally. A benefit of this architecture may be that the PNN 104B may affect the decision of the base DNN 104A during failure modes when it is activated. In addition, the PNN 104B may be smaller because it is not required to calculate each output of the base DNN 104A in addition to the new or alternative output. However, in a false positive scenario - for example, when the fusion mechanism 108 erroneously activates a patch - the result is the use of an output of the PNN 104B that does not include another instance of the earlier output of the base DNN 104A. This may result in less accurate results during a false positive scenario than in an embodiment where separate outputs 106 - for example, outputs 106A and 106B - are calculated (e.g., in an embodiment such as those of Figures 1B and 3B). Although shown as a single output 106, the base DNN / PNN combination may include separate output heads in embodiments such as those described herein.

[0038] Additionally, in some embodiments, a soft on / off switch may be used for PNN 104B. For example, instead of turning PNN 104B on or off, a gradual transition from no PNN 104B to use of PNN 104B may be applied. This may be achieved by multiplying the output of the patch node by a scaling factor, for example, from 0.0 (no application of PNN 104B) to 1.0 (full application of PNN 104B).

[0039] As another example, and with reference to FIG. 3B, the PNN 104B may be separate from the base DNN 104A such that the information processing of the base DNN 104A may not be changed from its initial state when the PNN 104B is not activated. For example, with reference to FIG. 1B, the added nodes of the PNN 104B affect the information processing of the base DNN 104A. As such, the architecture of FIG. 3B is similar to that of FIG. 1B, but allows selective activation of the PNN 104B based on the decision of the fusion mechanism 108. As a result, false positive scenarios may still include redundant outputs of the base DNN 104A, and false negative outputs will result in outputs similar to those of the base DNN 104A before patching. As such, computations may be reduced in instances where the PNN 104B is turned off.

[0040] In some embodiments, the PNN 104B and the fusion mechanism 108 can be tested and validated using real-world test data and / or simulated test data, e.g., in a simulation environment. For example, the PNN 104B and / or the fusion mechanism 108 can be tested in a resimulator, NVIDIA's DriveSIM, and / or in the real world. In some embodiments, after being tested and validated outside the consumer fleet, the PNN 104 and / or the fusion mechanism 108 (e.g., the gating NN 110) can be pushed over the air (OTA) to the consumer vehicle to run in a passive (inactive) mode, also known as a "shadow" mode. The passive mode can allow the base code to run, but the output 106 is not activated. As a result, the consumer fleet continues to operate, so that extensive testing can be performed in a short period of time. Events in which the fusion mechanism 108 decides to switch to the PNN 104B and / or events in which the outputs of the base DNN 104A and the PNN 104B do not match can be recorded and analyzed. Switching to the PNN 104B can be used to verify that the fusion mechanism is functioning correctly and that the PNN 104B correctly recovers from the failure mode of the base DNN 104A. The mismatch record can be used to verify that the PNN 104B operates correctly under normal conditions - for example, if the PNN 104B is selected incorrectly, the outputs 106B of the PNN 104B should still accurately reflect those of the outputs 106A of the base DNN 104A.

[0041] To implement the PNN 104B and the fusion mechanism 108 in addition to the base DNN 104A, the required computations of the PNN 104B and the fusion mechanism 108 may be pre-allocated - for example, sufficient buffer space may remain unused in the vehicle 500 upon initial deployment of the base DNN 104A to support the PNN 104B and the fusion mechanism 108. For the PNN 104B, this may mean that the maximum number of weights and nodes available for all combined PNNs 104B may be pre-defined. As such, this pre-allocated amount may be factored into the design decision of the PNN 104B. To pre-allocate the computations of the fusion mechanism 108, an estimate of the maximum CPU, GPU, DPU, accelerator, and / or other on-chip resources required may be determined. As such, both the PNN 104B and the fusion mechanism 108 may be added as "dummy" or placeholders to the computation graph to avoid scheduler changes since the changes may require additional validation work.

[0042] In some embodiments, a patch versioning system may be implemented to track the deployed PNN 104B, such as who requested the PNN 104B, who trained the PNN 104B and when, who validated the PNN 104B, and who, when, and to which vehicles the PNN 104B was deployed. Additionally, at least in part, simultaneously with and / or after the training and deployment of the PNN 104B, an updated base DNN 104A may be trained, tested, and validated to be pushed to the vehicles 500, for example, during the next major DNN release. As such, the PNN 104B and fusion mechanism 108 may function as a stop-gap measure during the training, testing, and validation period (e.g., in some instances, six months to a year) of the updated base DNN 104A.

[0043] Referring now to FIG. 4, each block of the method 400 described herein includes computational processes that may be performed using any combination of hardware, firmware, and / or software. For example, various functions may be performed by a processor executing instructions stored in a memory. The method 400 may also be implemented as computer usable instructions stored on a computer storage medium. The method 400 may be provided by a standalone application, a service, or a hosted service (standalone or in combination with another hosted service), or a plug-in to another product, to name a few. Additionally, the method 400 is described with respect to the process 100 of FIG. 1 and the vehicle 500 of FIGS. 5A-5D, as examples. However, the method 400 may additionally or alternatively be performed within any one system or any combination of systems, including but not limited to those described herein.

[0044] 4 is a flow diagram illustrating a method 400 for executing a PNN 104B during deployment, according to some embodiments of the present disclosure. The method 400 includes applying first data to a deep neural network (DNN) and a patch neural network (PNN) at block B402. For example, the input data 102 may be applied to the base DNN 104A and the PNN 104B.

[0045] The method 400 includes, in block B404, processing the first data to calculate the second data, which includes: (1) the DNN processes the first output of one or more nodes of the DNN; and (2) the PNN processes at least one first output of the second output of one or more nodes of the PNN and the one or more first outputs of the one or more nodes of the DNN. For example, the node of the layer 126A of the base DNN 104A may process the output of another node of the layer 126A of the base DNN 104A, and the PNN 104B may process both the output of the node of the layer 126A of the base DNN 104A and the output of the node of the layer 126B of the PNN 104B. Finally, the base DNN 104A may calculate the output 106A, and the PNN 104B may calculate the output 106B. However, in an embodiment, the output may be a single combined output in response to activation or deactivation of the PNN 104B, as described herein with respect to at least FIG. 3A. In this manner, the base DNN 104A may operate similarly to before patching, and the PNN 104B may address failure scenarios of the base DNN 104A.

[0046] The method 400 includes performing one or more actions based at least in part on the second data at block B406. For example, the output 106 may be used to perform one or more actions by the vehicle 500, such as object detection, classification, and / or tracking, path planning, control decisions, obstacle avoidance, and / or the like.

[0047] As a result, the PNN 104B can be implemented with a fast turnaround time to address failure scenarios of the base DNN 104A without requiring a full revalidation of the base DNN 104A. This can allow the vehicle 500 - or other underlying system - to continue operating while the updated base DNN 104A is trained and validated to account for the failure mode. In addition, the functionality of the base DNN 104A can remain unchanged when used with the PNN 104B, and the validated base DNN 104A can still be relied upon in all instances that do not address the failure mode, so that risk can be minimized in safety-critical applications - for example, autonomous or semi-autonomous driving.

[0048] Exemplary Autonomous Vehicle 5A is a diagram of an example autonomous vehicle 500 according to some embodiments of the present disclosure. The autonomous vehicle 500 (alternatively referred to herein as “vehicle 500”) may include, but is not limited to, a passenger vehicle, such as a car, a truck, a bus, a first responder vehicle, a shuttle, an electric or moped, a motorcycle, a fire engine, a police vehicle, an ambulance, a boat, a construction vehicle, a submarine, a drone, a vehicle coupled to a trailer, and / or another type of vehicle (e.g., unmanned and / or carrying one or more passengers). Autonomous vehicles are generally described in terms of automation levels as defined by the National Highway Traffic Safety Administration (NHTSA), a division of the U.S. Department of Transportation, and the Society of Automotive Engineers (SAE) "Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicle" (Standard No. J3016-201806 published June 15, 2018, Standard No. J3016-201609 published September 30, 2016, and previous and future versions of this standard). Vehicle 500 may be capable of functioning according to one or more of autonomous driving levels 3 through 5. For example, vehicle 500 may be capable of conditional automation (Level 3), highly automated (Level 4), and / or fully automated (Level 5), depending on the embodiment.

[0049] The vehicle 500 may include components such as a vehicle chassis, body, wheels (e.g., 2, 4, 6, 8, 18, etc.), tires, axles, and other components. The vehicle 500 may include a propulsion system 550, such as an internal combustion engine, a hybrid power plant, a fully electric engine, and / or another propulsion system type. The propulsion system 550 may be connected to a drive train of the vehicle 500, which may include a transmission, to enable propulsion of the vehicle 500. The propulsion system 550 may be controlled in response to receiving a signal from a throttle / accelerator 552.

[0050] Steering system 554, which may include a steering wheel, may be used to steer vehicle 500 (e.g., along a desired course or route) when propulsion system 550 is operating (e.g., when the vehicle is moving). Steering system 554 may receive signals from steering actuator 556. The steering wheel may be an option for fully automated (Level 5) functionality.

[0051] Brake sensor system 546 may be used to operate the vehicle brakes in response to receiving a signal from a brake actuator 548 and / or a brake sensor.

[0052] A controller 536, which may include one or more system on chip (SoC) 504 (FIG. 5C) and / or a GPU, may provide signals (e.g., representations of commands) to one or more components and / or systems of the vehicle 500. For example, the controller may send signals to operate vehicle brakes via one or more brake actuators 548, operate a steering system 554 via one or more steering actuators 556, and operate a propulsion system 550 via one or more throttle / accelerators 552. The controller 536 may include one or more on-board (e.g., integrated) computing devices (e.g., supercomputers) that process sensor signals and output operational commands (e.g., signals representing commands) to enable autonomous driving and / or assist a driver in driving the vehicle 500. The controllers 536 may include a first controller 536 for autonomous driving functions, a second controller 536 for functional safety functions, a third controller 536 for artificial intelligence functions (e.g., computer vision), a fourth controller 536 for infotainment functions, a fifth controller 536 for redundancy in emergency situations, and / or other controllers. In some instances, a single controller 536 may handle two or more of the aforementioned functions, and two or more controllers 536 may handle a single function, and / or any combination thereof.

[0053] The controller 536 may provide signals to control one or more components and / or systems of the vehicle 500 in response to sensor data (e.g., sensor inputs) received from one or more sensors. The sensor data may be received from, for example, and without limitation, global navigation satellite system sensors 558 (e.g., global positioning system sensors), RADAR sensors 560, ultrasonic sensors 562, LIDAR sensors 564, inertial measurement unit (IMU) sensors 566 (e.g., accelerometers, gyroscopes, magnetic compasses, magnetometers, etc.), microphones 596, stereo cameras 568, wide-view cameras 570 (e.g., fisheye cameras), infrared cameras 572, surround cameras 574 (e.g., 360 degree cameras), long-range and / or medium-range cameras 598, speed sensors 544 (e.g., for measuring the speed of the vehicle 500), vibration sensors 542, steering sensors 540, brake sensors (e.g., as part of a brake sensor system 546), and / or other sensor types.

[0054] One or more of the controllers 536 may receive input (e.g., represented by input data) from the instrument cluster 532 of the vehicle 500 and provide output (e.g., represented by output data, display data, etc.) via a human-machine interface (HMI) display 534, an audible annunciator, a loudspeaker, and / or other components of the vehicle 500. The output may include information such as vehicle velocity, speed, time, map data (e.g., HD map 522 of FIG. 5C), position data (e.g., the position of the vehicle 500 on a map, etc.), direction, the positions of other vehicles (e.g., occupancy grid), information regarding objects and object situations as known by the controller 536, etc. For example, the HMI display 534 may display information regarding the presence of one or more objects (e.g., road signs, warning signs, changing traffic signals, etc.) and / or a driving maneuver that the vehicle has performed, is performing, or will perform (e.g., changing lanes now, taking exit 34B in 2 miles, etc.).

[0055] Vehicle 500 further includes a network interface 524 capable of communicating over one or more networks using one or more wireless antennas 526 and / or a modem. For example, network interface 524 may be capable of communicating over LTE, WCDMA, UMTS, GSM, CDMA2000, etc. Wireless antenna 526 may also enable communication between objects (e.g., vehicles, mobile devices, etc.) in the environment using local area networks such as Bluetooth, Bluetooth LE, Z-Wave, ZigBee, etc., and / or low power wide-area networks (LPWANs) such as LoRaWAN, SigFox, etc.

[0056] 5B is an illustration of camera positions and fields of view of the example autonomous vehicle 500 of FIG. 5A, according to some embodiments of the present disclosure. The cameras and their respective fields of view are one example embodiment and are not intended to be limiting. For example, additional and / or alternative cameras may be included and / or the cameras may be located in different locations on the vehicle 500.

[0057] The camera type of the camera may include, but is not limited to, a digital camera that may be adapted for use with components and / or systems of the vehicle 500. The camera may be capable of operating at automotive safety integrity level (ASIL) B and / or at another ASIL. The camera type may be capable of any image capture rate, such as 60 frames per second (fps), 120 fps, 240 fps, etc., depending on the embodiment. The camera may be capable of using a rolling shutter, a global shutter, another type of shutter, or a combination thereof. In some instances, the color filter array may include a red clear clear clear (RCCC) color filter array, a red clear clear blue (RCCB) color filter array, a red blue green clear (RBGC) color filter array, a Foveon X3 color filter array, a Bayer sensor (RGGB) color filter array, a monochrome sensor color filter array, and / or another type of color filter array. In some embodiments, clear pixel cameras, such as cameras with RCCC, RCCB, and / or RBGC color filter arrays, may be used in an effort to increase light sensitivity.

[0058] In some instances, one or more of the cameras may be used to perform advanced driver assistance system (ADAS) functions (e.g., as part of a redundant or fail-safe design). For example, a multi-function mono camera may be installed to provide functions including lane departure warning, traffic sign assist, and intelligent headlamp control. One or more of the cameras (e.g., all of the cameras) may simultaneously record and provide image data (e.g., video).

[0059] One or more of the cameras may be mounted in a mounting part, such as a custom designed (3D printed) part, to filter out stray light and reflections from within the vehicle (e.g., reflections from the dashboard reflected in the windshield mirror) that may interfere with the camera's image data capture capabilities. With reference to a side mirror mounting part, the side mirror part may be custom 3D printed such that the camera mounting plate fits the shape of the side mirror. In some instances, the camera may be integrated into the side mirror. For side view cameras, the cameras may also be integrated into four posts at each corner of the cabin.

[0060] A camera (e.g., a forward-facing camera) with a field of view that includes a portion of the environment in front of the vehicle 500 may be used for surround view to help identify the forward path and obstacles and, with the aid of one or more controllers 536 and / or control SoCs, provide information essential to generating an occupancy grid and / or determining a preferred vehicle path. Forward-facing cameras may be used to perform many of the same ADAS functions as LIDAR, including emergency braking, pedestrian detection, and collision avoidance. Forward-facing cameras may also be used for ADAS functions and systems, including other functions such as lane departure warning ("LDW"), autonomous cruise control ("ACC"), and / or traffic sign recognition.

[0061] A variety of cameras may be used in a forward-facing configuration, including, for example, a monocular camera platform including a complementary metal oxide semiconductor (CMOS) color imager. Another example may be a wide-view camera 570 that may be used to capture objects entering the view from the periphery (e.g., pedestrians, crossing traffic, or bicycles). Although only one wide-view camera is shown in FIG. 5B, there may be any number of wide-view cameras 570 in the vehicle 500. In addition, a long-range camera 598 (e.g., a long-view stereo camera pair) may be used for depth-based object detection, especially for objects for which a neural network has not yet been trained. The long-range camera 598 may also be used for object detection and classification, as well as basic object tracking.

[0062] One or more stereo cameras 568 may also be included in the forward-facing configuration. The stereo camera 568 may include an integrated control unit with an extensible processing unit that may provide programmable logic (FPGA) and a multi-core microprocessor with CAN or Ethernet interfaces integrated on a single chip. Such a unit may be used to generate a 3D map of the vehicle's environment, including distance estimates for all points in the image. An alternative stereo camera 568 may include a compact stereo vision sensor that may include two camera lenses (one on the left and one on the right) and an image processing chip that can measure distances from the vehicle to objects of interest and use the generated information (e.g., metadata) to activate autonomous emergency braking and lane departure warning functions. Other types of stereo cameras 568 may be used in addition to or instead of those described herein.

[0063] A camera having a field of view that includes a portion of the environment to the side of the vehicle 500 (e.g., a side-view camera) may be used for surround view, providing information used to create and update the occupancy grid and to generate side impact collision warnings. For example, surround cameras 574 (e.g., four surround cameras 574 as shown in FIG. 5B) may be positioned on the vehicle 500. The surround cameras 574 may include a wide-view camera 570, a fisheye camera, a 360-degree camera, and / or the like. For example, four fisheye cameras may be positioned at the front, rear, and sides of the vehicle. In an alternative arrangement, the vehicle may use three surround cameras 574 (e.g., left, right, and rear) and may utilize one or more other cameras (e.g., a forward-facing camera) as a fourth surround view camera.

[0064] A camera having a field of view that includes the portion of the environment behind the vehicle 500 (e.g., a rear-view camera) may be used for parking assistance, surround view, rear collision warning, and creating and updating an occupancy grid. As described herein, a wide variety of cameras may be used, including, but not limited to, cameras that are also suitable as forward-facing cameras (e.g., long-range and / or mid-range camera 598, stereo camera 568, infrared camera 572, etc.).

[0065] FIG. 5C is a block diagram of an example system architecture of the example autonomous vehicle 500 of FIG. 5A, according to some embodiments of the present disclosure. It should be understood that this and other arrangements described herein are merely illustrative. Other arrangements and elements (e.g., machines, interfaces, functions, sequences, groupings of functions, etc.) may be used in addition to or instead of those shown, and some elements may be omitted altogether. Furthermore, many of the elements described herein are functional entities that may be implemented as separate or distributed components or in conjunction with other components, and in any suitable combination and location. Various functions described herein as being performed by entities may be implemented by hardware, firmware, and / or software. For example, various functions may be implemented by a processor executing instructions stored in a memory.

[0066] Each of the components, features, and systems of the vehicle 500 of FIG. 5C are shown as being connected via a bus 502. The bus 502 may include a controller area network (CAN) data interface (alternatively referred to as a “CAN bus”). The CAN may be a network within the vehicle 500 used to help control various features and functions of the vehicle 500, such as actuation of brakes, acceleration, braking, steering, windshield wipers, etc. The CAN bus may be configured to have dozens or even hundreds of nodes, each with its own unique identifier (e.g., CAN ID). The CAN bus may be read to find steering angle, ground speed, engine revolutions per minute (RPM), button positions, and / or other vehicle status indicators. The CAN bus may be ASIL B compliant.

[0067] The bus 502 is described herein as being a CAN bus, although this is not intended to be limiting. For example, FlexRay and / or Ethernet may be used in addition to or as an alternative to a CAN bus. Additionally, a single line is used to represent the bus 502, although this is not intended to be limiting. There may be any number of buses 502, which may include, for example, one or more CAN buses, one or more FlexRay buses, one or more Ethernet buses, and / or one or more other types of buses using different protocols. In some instances, two or more buses 502 may be used to perform different functions and / or may be used for redundancy. For example, a first bus 502 may be used for collision avoidance functions and a second bus 502 may be used for operational control. In any instance, each bus 502 may communicate with any of the components of the vehicle 500, and two or more buses 502 may communicate with the same component. In some examples, each SoC 504, each controller 536, and / or each computer in the vehicle may have access to the same input data (e.g., input from sensors in the vehicle 500) and may be connected to a common bus, such as a CAN bus.

[0068] 5A. The controller 536 may be used for a variety of functions. The controller 536 may be coupled to any of a variety of other components and systems of the vehicle 500 and may be used for control of the vehicle 500, artificial intelligence of the vehicle 500, infotainment for the vehicle 500, and / or the like.

[0069] The vehicle 500 may include a system on a chip (SoC) 504. The SoC 504 may include a CPU 506, a GPU 508, a processor 510, a cache 512, an accelerator 514, a data store 516, and / or other components and features not shown. The SoC 504 may be used to control the vehicle 500 in a variety of platforms and systems. For example, the SoC 504 may be combined in a system (e.g., the system of the vehicle 500) having an HD map 522 that can obtain map refreshes and / or updates via a network interface 524 from one or more servers (e.g., server 578 of FIG. 5D ).

[0070] CPU 506 may include a CPU cluster or CPU complex (alternatively referred to as a "CCPLEX"). CPU 506 may include multiple cores and / or L2 caches. For example, in some embodiments, CPU 506 may include eight cores in a coherent multiprocessor configuration. In some embodiments, CPU 506 may include four dual-core clusters, each with its own dedicated L2 cache (e.g., a 2MBL2 cache). CPU 506 (e.g., CCPLEX) may be configured to support concurrent cluster operation allowing any combination of CPU 506 clusters to be active at any given time.

[0071] The CPU 506 may implement power management capabilities including one or more of the following features: individual hardware blocks may be automatically clock gated when idle to conserve dynamic power, each core clock may be gated when the core is not actively executing instructions by execution of WFI / WFE instructions, each core may be independently power gated, each core cluster may be independently clock gated when all cores are clock gated or power gated, and / or each core cluster may be independently power gated when all cores are power gated. The CPU 506 may further implement an enhanced algorithm for managing power states where allowed power states and expected wake-up times are specified and the hardware / microcode determines the best power state for entering the cores, clusters, and CCPLEXes. The processing cores may support simplified power state entry sequences in software where work is offloaded to the microcode.

[0072] The GPU 508 may include an integrated GPU (alternatively referred to herein as an "iGPU"). The GPU 508 may be programmable and efficient for parallel workloads. In some instances, the GPU 508 may use an enhanced tensor instruction set. The GPU 508 may include one or more streaming microprocessors, where each streaming microprocessor may include an L1 cache (e.g., an L1 cache having at least 96 KB storage capacity) and two or more of the streaming microprocessors may share a cache (e.g., an L2 cache having 512 KB storage capacity). In some embodiments, the GPU 508 may include at least eight streaming microprocessors. The GPU 508 may use a computational application programming interface (API). In addition, the GPU 508 may use one or more parallel computing platforms and / or programming models (e.g., NVIDIA's CUDA).

[0073] The GPU 508 may be power optimized for best performance in automotive and embedded use cases. For example, the GPU 508 may be fabricated on FinFETs (Fin field-effect transistors). However, this is not intended to be limiting and the GPU 508 may be fabricated using other semiconductor fabrication processes. Each streaming microprocessor may incorporate several mixed-precision processing cores partitioned into multiple blocks. For example, and without limitation, 64 PF32 cores and 32 PF64 cores may be partitioned into four processing blocks. In such an example, each processing block may be assigned 16 FP32 cores, 8 FP64 cores, 16 INT32 cores, two mixed-precision NVIDIA tensor cores for deep learning matrix operations, an L0 instruction cache, a warp scheduler, a dispatch unit, and / or a 64KB register file. In addition, the streaming microprocessor may include independent parallel integer and floating-point data paths to provide efficient execution of workloads having a mix of computational and addressing operations. Streaming microprocessors may include independent thread scheduling capabilities to allow finer grain synchronization and coordination among concurrent threads. Streaming microprocessors may include a combined L1 data cache and shared memory unit to improve performance while simplifying programming.

[0074] The GPU 508 may include high bandwidth memory (HBM) and / or a 16GB HBM2 memory subsystem to provide about 900GB / sec peak memory bandwidth in some instances. In some instances, synchronous graphics random-access memory (SGRAM), such as graphics double data rate type five synchronous random-access memory (GDDR5), may be used in addition to or in place of the HBM memory.

[0075] The GPU 508 may include unified memory technology including access counters to enable more accurate movement of memory pages to the processors that access them most frequently, thereby improving the efficiency of storage ranges shared between processors. In some instances, address translation service (ATS) support may be used to enable the GPU 508 to directly access the CPU 506 page tables. In such instances, when the GPU 508 memory management unit (MMU) experiences a miss, an address translation request may be sent to the CPU 506. In response, the CPU 506 may consult its page tables for a virtual-to-real mapping of the address and send the translation back to the GPU 508. As such, the unified memory technology may enable a single unified virtual address space for both the CPU 506 and GPU 508 memories, thereby simplifying GPU 508 programming and porting of applications to the GPU 508.

[0076] In addition, GPU 508 may include access counters that can record the frequency of accesses of GPU 508 to the memory of other processors. The access counters can help ensure that memory pages are moved to the physical memory of the processors that are accessing the pages most frequently.

[0077] The SoC 504 may include any number of caches 512, including those described herein. For example, the cache 512 may include an L3 cache available to both the CPU 506 and the GPU 508 (e.g., connected to both the CPU 506 and the GPU 508). The cache 512 may include a write-back cache that can record line state, such as by using a cache coherence protocol (e.g., MEI, MESI, MSI, etc.). The L3 cache may include 4 MB or more, depending on the embodiment, although smaller cache sizes may be used.

[0078] The SoC 504 may include an arithmetic logic unit (ALU) that may be utilized in performing processing for any of the various tasks or operations (e.g., processing DNNs) of the vehicle 500. In addition, the SoC 504 may include a floating point unit (FPU) (or other math co-processor or numeric co-processor type) for performing mathematical operations within the system. For example, the SoC 104 may include one or more FPUs integrated as execution units within the CPU 506 and / or GPU 508.

[0079] The SoC 504 may include one or more accelerators 514 (e.g., hardware accelerators, software accelerators, or a combination thereof). For example, the SoC 504 may include a hardware acceleration cluster, which may include optimized hardware accelerators and / or large on-chip memory. The large on-chip memory (e.g., 4MB of SRAM) may enable the hardware acceleration cluster to accelerate neural networks and other operations. The hardware acceleration cluster may be used to complement the GPU 508 and to offload some of the tasks of the GPU 508 (e.g., to free up more cycles of the GPU 508 to perform other tasks). As an example, the accelerator 514 may be used for target workloads that are stable enough to be suitable for acceleration (e.g., perception, convolutional neural networks (CNNs), etc.). As used herein, the term "CNN" may include all types of CNNs, including region-based or regional convolutional neural networks (RCNNs) and Faster RCNNs (e.g., as used for object detection).

[0080] The accelerator 514 (e.g., a hardware acceleration cluster) may include a deep learning accelerator (DLA). The DLA may include one or more tensor processing units (TPUs) that may be configured to provide an additional 10 trillion operations per second for deep learning applications and inference. The TPU may be an accelerator configured and optimized to perform image processing functions (e.g., CNN, RCNN, etc.). The DLA may further be optimized for a specific set of neural network types and floating point operations, as well as inference. The design of the DLA may provide more performance per millimeter than a general-purpose GPU, greatly exceeding the performance of a CPU. The TPU may perform several functions, including, for example, single instance convolution functions, supporting INT8, INT16, and FP16 data types for both features and weights, as well as post-processor functions.

[0081] The DLA can quickly and efficiently run neural networks, in particular CNNs, on processed or unprocessed data for any of a variety of functions, including but not limited to: CNNs for object identification and detection using data from camera sensors, CNNs for distance estimation using data from camera sensors, CNNs for emergency vehicle detection and identification using data from microphones, CNNs for face recognition and vehicle owner identification using data from camera sensors, and / or CNNs for security and / or safety related events.

[0082] The DLA can perform any function of the GPU 508, and by using an inference accelerator, for example, a designer can target either the DLA or the GPU 508 for any function. For example, a designer can focus on processing CNNs and floating point operations on the DLA, and offload other functions to the GPU 508 and / or other accelerators 514.

[0083] The accelerator 514 (e.g., hardware acceleration cluster) may include a programmable vision accelerator (PVA), which may alternatively be referred to herein as a computer vision accelerator. The PVA may be designed and configured to accelerate computer vision algorithms for advanced driver assistance systems (ADAS), autonomous driving, and / or augmented reality (AR) and / or virtual reality (VR) applications. The PVA may provide a balance between performance and flexibility. For example, each PVA may include, but is not limited to, any number of reduced instruction set computer (RISC) cores, direct memory access (DMA), and / or any number of vector processors, for example.

[0084] The RISC cores may interact with an image sensor (e.g., an image sensor of any of the cameras described herein), an image signal processor, and / or the like. Each RISC core may include any amount of memory. The RISC cores may use any of a number of protocols, depending on the embodiment. In some instances, the RISC cores may run a real-time operating system (RTOS). The RISC cores may be implemented using one or more integrated circuit devices, application specific integrated circuits (ASICs), and / or memory devices. For example, the RISC cores may include an instruction cache and / or tightly coupled RAM.

[0085] The DMA may allow components of the PVA to access system memory independent of the CPU 506. The DMA may support any number of features used to provide optimizations to the PVA, including but not limited to supporting multi-dimensional addressing and / or circular addressing. In some instances, the DMA may support up to six or more dimensions of addressing, which may include block width, block height, block depth, horizontal block stepping, vertical block stepping, and / or depth stepping.

[0086] The vector processor may be a programmable processor that may be designed to efficiently and flexibly execute computer vision algorithm programming and provide signal processing capabilities. In some instances, the PVA may include a PVA core and two vector processing subsystem partitions. The PVA core may include a processor subsystem, a DMA engine (e.g., two DMA engines), and / or other peripherals. The vector processing subsystem may act as the primary processing engine of the PVA and may include a vector processing unit (VPU), an instruction cache, and / or a vector memory (e.g., VMEM). The VPU core may include a digital signal processor, such as, for example, a single instruction, multiple data (SIMD), very long instruction word (VLIW) digital signal processor. The combination of SIMD and VLIW may increase throughput and speed.

[0087] Each vector processor may include an instruction cache and may be coupled to dedicated memory. As a result, in some instances, each vector processor may be configured to execute independently of other vector processors. In other instances, the vector processors included in a particular PVA may be configured to employ data parallelism. For example, in some embodiments, multiple vector processors included in a single PVA may execute the same computer vision algorithm, but on different regions of an image. In other instances, the vector processors included in a particular PVA may execute different computer vision algorithms simultaneously on the same image, or even execute different algorithms on sequential images or portions of an image. In particular, any number of PVAs may be included in a hardware accelerated cluster, and any number of vector processors may be included in each PVA. In addition, the PVAs may include additional error correcting code (ECC) memory to increase overall system security.

[0088] The accelerator 514 (e.g., a hardware acceleration cluster) may include a computer vision network-on-chip and SRAM to provide high bandwidth, low latency SRAM for the accelerator 514. In some instances, the on-chip memory may include, for example and without limitation, at least 4 MB of SRAM consisting of eight field configurable memory blocks that may be accessible by both the PVA and DLA. Each pair of memory blocks may include an advanced peripheral bus (APB) interface, configuration circuitry, controllers, and multiplexers. Any type of memory may be used. The PVA and DLA may access the memory through a backbone that provides the PVA and DLA with high speed access to the memory. The backbone may include a computer vision network-on-chip that interconnects the PVA and DLA to the memory (e.g., using the APB).

[0089] The computer vision network-on-chip may include an interface that determines, prior to transmission of any control signals / addresses / data, that both the PVA and DLA provide ready and valid signals. Such an interface may provide separate phases and separate channels for transmitting control signals / addresses / data, as well as burst type communication for continuous data transfer. This type of interface may follow ISO26262 or IEC61508 standards, although other standards and protocols may be used.

[0090] In some instances, the SoC 504 may include a real-time ray tracing hardware accelerator, such as that described in U.S. Pat. No. 6,399,633, which may be used to quickly and efficiently determine the location and scale of objects (e.g., within a world model) to generate real-time visualization simulations for RADAR signal interpretation, for acoustic propagation synthesis and / or analysis, for simulation of SONAR systems, for general wave propagation simulations, for comparison against LIDAR data for localization and / or other functions, and / or for other uses. In some examples, one or more tree traversal units (TTUs) may be used to perform one or more ray tracing related operations.

[0091] The accelerator 514 (e.g., a hardware accelerator cluster) has a variety of applications for autonomous driving. The PVA may be a programmable vision accelerator that can be used for critical processing stages in ADAS and autonomous vehicles. The capabilities of the PVA are suited to the domain of algorithms that require predictable processing at low power and low latency. In other words, the PVA works well with semi-dense or dense regular computations on small data sets that require predictable execution times with low latency and low power. Hence, in the context of a platform for autonomous vehicles, the PVA is designed to run classical computer vision algorithms, since the PVA is efficient in object detection and operating on integer computations.

[0092] For example, according to one embodiment of the present technology, the PVA is used to perform computer stereo vision. A semi-global matching based algorithm may be used in some instances, but this is not intended to be limiting. Many applications for Level 3-5 autonomous driving require motion estimation / stereo matching on the fly (e.g., structure from motion, pedestrian recognition, lane detection, etc.). The PVA can perform computer stereo vision functions with input from two monocular cameras.

[0093] In some instances, PVA may be used to perform dense optical flow by processing raw RADAR data (e.g., using a 4D Fast Fourier Transform) to provide processed RADAR. In other instances, PVA is used for time of flight depth processing, for example, by processing raw time of flight data to provide processed time of flight data.

[0094] DLA can be used to implement any type of network to enhance control and driving safety, including, for example, a neural network that outputs a measure of the confidence of each object detection. Such confidence value can be interpreted as a probability or as providing a relative "weight" of each detection compared to other detections. This confidence value allows the system to make further decisions regarding which detections should be considered true positive detections rather than false positive detections. For example, the system can set a confidence threshold and consider only detections above the threshold as true positive detections. In an automatic emergency braking (AEB) system, a false positive detection would cause the vehicle to automatically perform emergency braking, which is clearly undesirable. Therefore, only the most confident detections should be considered as triggers for AEB. DLA can implement a neural network that regresses the confidence value. The neural network may receive as its inputs at least some subset of parameters, such as bounding box dimensions, ground plane estimates obtained (e.g., from another subsystem), inertial measurement unit (IMU) sensor 566 outputs that correlate with vehicle 500 orientation, range, and 3D position estimates of objects obtained from the neural network and / or other sensors (e.g., LIDAR sensor 564 or RADAR sensor 560), and others.

[0095] The SoC 504 may include a data store 516 (e.g., memory). The data store 516 may be an on-chip memory of the SoC 504 and may store the neural network to be executed by the GPU and / or DLA. In some instances, the data store 516 may have a capacity large enough to store multiple instances of the neural network for redundancy and safety. The data store 512 may comprise an L2 or L3 cache 512. References to the data store 516 may include references to memory associated with the PVA, DLA, and / or other accelerators 514, as described herein.

[0096] The SoC 504 may include one or more processors 510 (e.g., embedded processors). The processors 510 may include a boot and power management processor, which may be a dedicated processor and subsystem for handling boot power and management capabilities and associated security enforcement. The boot and power management processor may be part of the SoC 504 boot sequence and may provide run-time power management services. The boot power and management processor may provide clock and voltage programming, assist with system low power state transitions, manage the SoC 504 thermal and temperature sensors, and / or manage the SoC 504 power states. Each temperature sensor may be implemented as a ring oscillator whose output frequency is proportional to temperature, and the SoC 504 may use the ring oscillator to detect the temperature of the CPU 506, the GPU 508, and / or the accelerator 514. If the temperature is determined to exceed a threshold, the boot and power management processor may enter a temperature fault routine, place the SoC 504 in a lower power state and / or place the vehicle 500 in a Chauffeur safe shutdown mode (e.g., bring the vehicle 500 to a safe shutdown).

[0097] The processor 510 may further include a set of embedded processors that may perform the functions of an audio processing engine. The audio processing engine may be an audio subsystem that allows full hardware support of multi-channel audio through multiple interfaces and a wide and flexible range of audio I / O interfaces. In some instances, the audio processing engine is a dedicated processor core that includes a digital signal processor with dedicated RAM.

[0098] The processor 510 may further include an always-on processor engine that can provide the necessary hardware features to support low power sensor management and wake use cases. The always-on processor engine may include a processor core, tightly coupled RAM, support peripherals (e.g., timers and interrupt controllers), various I / O controller peripherals, and routing logic.

[0099] The processor 510 may further include a safety cluster engine that includes a processor subsystem dedicated to handling safety management of automotive applications. The safety cluster engine may include two or more processor cores, tightly coupled RAM, supporting peripherals (e.g., timers, interrupt controllers, etc.), and / or routing logic. In a safety mode, the two or more cores may operate in lockstep mode and function as a single core with comparison logic to detect any differences between their operations.

[0100] The processor 510 may further include a real-time camera engine, which may include a dedicated processor subsystem for handling real-time camera management.

[0101] The processor 510 may further include a high dynamic range signal processor, which may include an image signal processor, which is a hardware engine that is part of the camera processing pipeline.

[0102] The processor 510 may include a video image compositor, which may be a processing block (e.g., implemented in a microprocessor) that implements video post-processing functions required by the video playback application to produce a final image for the player window. The video image compositor may perform lens distortion correction on the wide-view camera 570, on the surround camera 574, and / or on the in-cabin surveillance camera sensor. The in-cabin surveillance camera sensor is preferably monitored by a neural network running on another instance of the advanced SoC, configured to identify in-cabin events and respond appropriately. The in-cabin system may perform lip reading to activate cellular service and make phone calls, dictate emails, change the vehicle's destination, activate or change the vehicle's infotainment system and settings, or provide voice-activated web surfing. Certain features are only available to the driver when operating in autonomous mode and are disabled otherwise.

[0103] The video image synthesizer may include enhanced temporal noise reduction for both spatial and temporal noise reduction. For example, when motion occurs in the video, the noise reduction reduces the weight of information provided by adjacent frames and appropriately weights the spatial information. When an image or part of an image does not contain motion, the temporal noise reduction performed by the video image synthesizer may use information from previous images to reduce noise in the current image.

[0104] The video image composer may also be configured to perform stereo rectification on the input stereo lens frames. The video image composer may further be used for user interface compositing when the operating system desktop is in use and the GPU 508 is not required to continuously render new surfaces. Even when the GPU 508 is powered on and actively performing 3D rendering, the video image composer may be used to offload the GPU 508 for improved performance and responsiveness.

[0105] The SoC504 may further include a mobile industry processor interface (MIPI) camera serial interface for receiving video and input from a camera, a high-speed interface, and / or a video input block that may be used for camera and related pixel input functions. The SoC504 may further include an input / output controller that may be controlled by software and that may be used to receive I / O signals that are not committed to a specific role.

[0106] The SoC 504 may further include a wide range of peripheral interfaces to enable peripherals, audio codecs, power management, and / or communication with other devices. The SoC 504 may be used to process data from cameras (e.g., connected via gigabit multimedia serial links and Ethernet), sensors (e.g., LIDAR sensor 564, RADAR sensor 560, etc., which may be connected via Ethernet), data from bus 502 (e.g., speed of vehicle 500, steering wheel position, etc.), and data from GNSS sensor 558 (e.g., connected via Ethernet or CAN bus). The SoC 504 may further include a dedicated high-performance mass storage controller, which may include its own DMA engine and may be used to offload the CPU 506 from routine data management tasks.

[0107] The SoC504 may be an end-to-end platform with a flexible architecture that spans levels 3-5 of automation, thereby providing a comprehensive functional safety architecture that leverages and efficiently uses computer vision and ADAS techniques for diversity and redundancy, and provides a platform for a flexible, reliable driving software stack along with deep learning tools. The SoC504 may be faster, more reliable, and more energy- and space-efficient than traditional systems. For example, when the accelerator 514 is combined with the CPU 506, the GPU 508, and the data store 516, the SoC504 may provide a fast and efficient platform for levels 3-5 of autonomous vehicles.

[0108] Thus, the present technology provides capabilities and functionality that cannot be achieved by conventional systems. For example, computer vision algorithms may be executed on a CPU, which may be configured using a high-level programming language, such as the C programming language, to execute a wide variety of processing algorithms across a wide variety of visual data. However, CPUs often cannot meet the performance requirements of many computer vision applications, such as those related to execution time and power consumption. In particular, many CPUs cannot execute complex object detection algorithms in real time, which is a requirement for in-vehicle ADAS applications and for practical Level 3-5 autonomous vehicles.

[0109] In contrast to conventional systems, by providing a CPU complex, a GPU complex, and a hardware acceleration cluster, the techniques described herein allow multiple neural networks to run simultaneously and / or sequentially and the results to be combined to enable Level 3-5 autonomous driving capabilities. For example, the CNN running on the DLA or dGPU (e.g., GPU520) can include text and word recognition that allows the supercomputer to read and understand traffic signs, including signs for which the neural network was not specifically trained. The DLA can further include neural networks that can identify, interpret, and provide a semantic understanding of the signs and pass the semantic understanding to a route planning module running on the CPU complex.

[0110] As another example, multiple neural networks may be run simultaneously, as required for level 3, 4, or 5 driving. For example, a warning sign consisting of "Caution: Flashing lights indicate icy conditions" along with a lightning flash may be interpreted by several neural networks, independently or collectively. The sign itself may be identified as a traffic sign by a first deployed neural network (e.g., a trained neural network), and the text "Flashing lights indicate icy conditions" may be interpreted by a second deployed neural network that informs the vehicle's route planning software (preferably running on the CPU complex) that icy conditions exist when a flashing light is detected. The flashing light may be identified by running a third deployed neural network through multiple frames, informing the vehicle's route planning software of the presence (or absence) of the flashing light. All three neural networks may be run simultaneously, such as within the DLA and / or on the GPU 508.

[0111] In some instances, a CNN for facial recognition and vehicle owner identification can use data from the camera sensors to identify the presence of a legitimate driver and / or owner of the vehicle 500. An always-on sensor processing engine can be used to unlock the vehicle and turn on the lights when the owner approaches the driver's door, and in security mode, to disable vehicle operation when the owner leaves the vehicle. In this manner, the SoC 504 provides security against theft and / or carjacking.

[0112] In another example, a CNN for emergency vehicle detection and identification can detect and identify emergency vehicle sirens using data from microphone 596. In contrast to conventional systems that use general classifiers to detect sirens and manually extract features, SoC 504 uses CNNs for classification of environmental and urban sounds, as well as classification of visual data. In a preferred embodiment, the CNN running on the DLA is trained to identify the relative terminal velocity of emergency vehicles (e.g., by using the Doppler effect). The CNN can also be trained to identify emergency vehicles specific to the local area in which the vehicle is operating, as identified by the GNSS sensor 558. Thus, for example, when operating in Europe, the CNN will attempt to detect European sirens, and when in the United States, the CNN will attempt to identify only North American sirens. After an emergency vehicle is detected, a control program can be used to perform emergency vehicle safety routines, such as slowing the vehicle down, stopping at the side of the road, parking the vehicle, and / or idling the vehicle, with the assistance of ultrasonic sensor 562, until the emergency vehicle has passed.

[0113] The vehicle may include a CPU 518 (e.g., a separate CPU, or a dCPU) that may be coupled to the SoC 504 via a high-speed interconnect (e.g., PCIe). The CPU 518 may include, for example, an X86 processor. The CPU 518 may be used to perform any of a variety of functions, including, for example, reconciling potentially inconsistent results between the ADAS sensors and the SoC 504, and / or monitoring the status and health of the controller 536 and / or the infotainment SoC 530.

[0114] Vehicle 500 may include a GPU 520 (e.g., a discrete GPU, or a dGPU) that may be coupled to SoC 504 via a high-speed interconnect (e.g., NVIDIA's NVLINK). GPU 520 may provide additional artificial intelligence capabilities, such as by running redundant and / or different neural networks, and may be used to train and / or update neural networks based on inputs (e.g., sensor data) from sensors of vehicle 500.

[0115] The vehicle 500 may further include a network interface 524, which may include one or more wireless antennas 526 (e.g., one or more wireless antennas for different communication protocols, such as a cellular antenna, a Bluetooth antenna, etc.). The network interface 524 may be used to enable wireless connections with the cloud over the Internet (e.g., with a server 578 and / or other network devices), with other vehicles, and / or with computing devices (e.g., passenger client devices). To communicate with other vehicles, a direct link may be established between the two vehicles and / or an indirect link may be established (e.g., through a network and via the Internet). A direct link may be provided using a vehicle-to-vehicle communication link. The vehicle-to-vehicle communication link may provide the vehicle 500 information about vehicles in close proximity to the vehicle 500 (e.g., vehicles in front of, beside, and / or behind the vehicle 500). This function may be part of a cooperative adaptive cruise control function of the vehicle 500.

[0116] The network interface 524 may include a SoC that provides modulation and demodulation functions and enables the controller 536 to communicate over a wireless network. The network interface 524 may include a radio frequency front end for upconversion from baseband to radio frequency and downconversion from radio frequency to baseband. The frequency conversion may be performed through well known processes and / or may be performed using a superheterodyne process. In some instances, the radio frequency front end functions may be provided by a separate chip. The network interface may include wireless functions for communicating via LTE, WCDMA, UMTS, GSM, CDMA2000, Bluetooth, Bluetooth LE, Wi-Fi, Z-Wave, ZigBee, LoRaWAN, and / or other wireless protocols.

[0117] The vehicle 500 may further include a data store 528, which may include off-chip (e.g., off-SoC 504) storage. The data store 528 may include one or more memory elements, including RAM, SRAM, DRAM, VRAM, flash, hard disk, and / or other components and / or devices capable of storing at least one bit of data.

[0118] The vehicle 500 may further include a GNSS sensor 558. The GNSS sensor 558 (e.g., a GPS, an aided GPS sensor, a differential GPS (DGPS) sensor, etc.) aids in mapping, perception, occupancy grid generation, and / or route planning functions. Any number of GNSS sensors 558 may be used, including, but not limited to, a GPS using a USB connector with an Ethernet to serial (RS-232) bridge, for example.

[0119] The vehicle 500 may further include a RADAR sensor 560. The RADAR sensor 560 may be used by the vehicle 500 for long-range vehicle detection, even in darkness and / or severe weather conditions. The RADAR functional safety level may be ASIL B. In some instances, the RADAR sensor 560 may use the CAN and / or bus 502 for control and to access object tracking data (e.g., to transmit data generated by the RADAR sensor 560), with access to Ethernet to access raw data. A wide variety of RADAR sensor types may be used. For example, and without limitation, the RADAR sensor 560 may be suitable for front, rear, and side RADAR use. In some instances, a pulsed Doppler RADAR sensor is used.

[0120] The RADAR sensor 560 may include different configurations, such as long range with a narrow field of view, short range with a wide field of view, and short range side coverage. In some instances, the long range RADAR may be used for adaptive cruise control functions. Long range RADAR systems may provide a wide field of view achieved by two or more independent scans, such as within a 250m range. The RADAR sensor 560 may help distinguish between static and moving objects and may be used by ADAS systems for emergency brake assist and forward collision warning. The long range RADAR sensor may include a monostatic multimodal RADAR with multiple (e.g., six or more) fixed RADAR antennas and high speed CAN and FlexRay interfaces. In one instance with six antennas, the center four antennas may create a focused beam pattern designed to record the surroundings of the vehicle 500 at high speeds with minimal interference from traffic in adjacent lanes. The other two antennas may increase the field of view, allowing for quick detection of vehicles entering or leaving the lane of the vehicle 500.

[0121] As an example, a mid-range RADAR system may include a range of up to 560m (front) or 80m (rear) and a field of view of up to 42 degrees (front) or 550 degrees (rear). A short-range RADAR system may include, but is not limited to, a RADAR sensor designed to be mounted on either end of the rear bumper. When mounted on either end of the rear bumper, such a RADAR sensor system can create two beams that constantly monitor the blind spots behind and next to the vehicle.

[0122] Short-range RADAR systems may be used in ADAS systems for blind spot detection and / or lane change assist.

[0123] The vehicle 500 may further include ultrasonic sensors 562. The ultrasonic sensors 562, which may be located at the front, rear, and / or sides of the vehicle 500, may be used for parking assist and / or for creating and updating an occupancy grid. A wide variety of ultrasonic sensors 562 may be used, with different ultrasonic sensors 562 being used for different ranges of detection (e.g., 2.5m, 4m). The ultrasonic sensors 562 may operate at a functional safety level of ASIL B.

[0124] The vehicle 500 may include a LIDAR sensor 564. The LIDAR sensor 564 may be used for object and pedestrian detection, emergency braking, collision avoidance, and / or other functions. The LIDAR sensor 564 may be functional safety level ASIL B. In some instances, the vehicle 500 may include multiple (e.g., 2, 4, 6, etc.) LIDAR sensors 564 that may use Ethernet (e.g., to provide data to a Gigabit Ethernet switch).

[0125] In some instances, the LIDAR sensor 564 may have the ability to provide a list of objects and their distances in a 360-degree field of view. A commercially available LIDAR sensor 564 may have an advertised range of about 500m, for example, with an accuracy of 2cm to 3cm, and with support for a 500Mbps Ethernet connection. In some instances, one or more non-protruding LIDAR sensors 564 may be used. In such instances, the LIDAR sensor 564 may be implemented as a small device that may be integrated into the front, rear, sides, and / or corners of the vehicle 500. In such instances, the LIDAR sensor 564 may have a range of 200m even for low-reflecting objects and provide up to 120-degree horizontal and 35-degree vertical fields of view. A front-mounted LIDAR sensor 564 may be configured for a horizontal field of view between 45 and 135 degrees.

[0126] In some instances, LIDAR technology such as 3D flash LIDAR may also be used. 3D flash LIDAR uses a laser flash as a transmitter to illuminate the surroundings of the vehicle up to about 200 m. The flash LIDAR unit includes a receptor that records the laser pulse transit time and the reflected light on each pixel, which in turn corresponds to the range from the vehicle to the object. Flash LIDAR may allow a highly accurate and undistorted image of the surroundings to be generated with every laser flash. In some instances, four flash LIDAR sensors may be deployed, one on each side of the vehicle 500. Available 3D flash LIDAR systems include solid-state 3D steering array LIDAR cameras that have no moving parts other than the blower (e.g., non-scanning LIDAR devices). Flash LIDAR devices may use 5 nanosecond Class I (eye-safe) laser pulses per frame and may capture reflected laser light in the form of a 3D range point cloud and coregistered intensity data. By using flash LIDAR, and because flash LIDAR is a solid-state device with no moving parts, the LIDAR sensor 564 may be less susceptible to motion blur, vibration, and / or shock.

[0127] The vehicle may further include an IMU sensor 566. In some instances, the IMU sensor 566 may be positioned at the center of the rear axle of the vehicle 500. The IMU sensor 566 may include, for example, but not limited to, an accelerometer, a magnetometer, a gyroscope, a magnetic compass, and / or other sensor types. In some instances, such as in a 6-axis application, the IMU sensor 566 may include an accelerometer and a gyroscope, while in a 9-axis application, the IMU sensor 566 may include an accelerometer, a gyroscope, and a magnetometer.

[0128] In some embodiments, the IMU sensor 566 may be implemented as a miniature, high-performance GPS-Aided Inertial Navigation System (GPS / INS) that combines micro-electro-mechanical system (MEMS) inertial sensors, highly sensitive GPS receivers, and advanced Kalman filtering algorithms to provide estimates of position, velocity, and attitude. As such, in some instances, the IMU sensor 566 may enable the vehicle 500 to estimate heading without requiring input from a magnetic sensor by directly observing and correlating changes in velocity from the GPS to the IMU sensor 566. In some instances, the IMU sensor 566 and the GNSS sensor 558 may be combined in a single integrated unit.

[0129] The vehicle may include microphones 596 placed in and / or around the vehicle 500. The microphones 596 may be used for emergency vehicle detection and identification, among other things.

[0130] The vehicle may further include any number of camera types, including stereo cameras 568, wide view cameras 570, infrared cameras 572, surround cameras 574, long and / or mid-range cameras 598, and / or other camera types. The cameras may be used to capture image data around the entire exterior of the vehicle 500. The type of cameras used depends on the embodiment and requirements of the vehicle 500, and any combination of camera types may be used to achieve the required coverage around the vehicle 500. Additionally, the number of cameras may vary depending on the embodiment. For example, the vehicle may include six cameras, seven cameras, ten cameras, twelve cameras, and / or another number of cameras. The cameras may support, by way of example only, Gigabit Multimedia Serial Link (GMSL) and / or Gigabit Ethernet. Each camera is described in further detail herein with respect to FIGS. 5A and 5B.

[0131] The vehicle 500 may further include a vibration sensor 542. The vibration sensor 542 may measure vibrations of vehicle components, such as an axle. For example, a change in vibration may indicate a change in the surface of the road. In another example, when two or more vibration sensors 542 are used, the difference in vibration may be used to determine friction or slippage of the road surface (e.g., when the difference in vibration is between a powered axle and a free-running axle).

[0132] The vehicle 500 may include an ADAS system 538. In some instances, the ADAS system 538 may include a SoC. The ADAS system 538 may include autonomous / adaptive / automatic cruise control (ACC), cooperative adaptive cruise control (CACC), forward crash warning (FCW), automatic emergency braking (AEB), lane departure warning (LDW), lane keep assist (LKA), blind spot warning (BSW), rear cross-traffic warning (RCTW), collision warning system (CWS), lane centering (LC), and / or other features and functions.

[0133] The ACC system may use a RADAR sensor 560, a LIDAR sensor 564, and / or a camera. The ACC system may include longitudinal ACC and / or lateral ACC. Longitudinal ACC monitors and controls the distance to the vehicle directly ahead of the vehicle 500 and automatically adjusts the vehicle speed to maintain a safe distance from the vehicle ahead. Lateral ACC performs distance keeping and advises the vehicle 500 to change lanes when necessary. Lateral ACC is related to other ADAS applications such as LCA and CWS.

[0134] CACC uses information from other vehicles, which may be received from other vehicles via a wireless link via the network interface 524 and / or the wireless antenna 526, or indirectly via a network connection (e.g., via the Internet). A direct link may be provided by a vehicle-to-vehicle (V2V) communication link, while an indirect link may be an infrastructure-to-vehicle (I2V) communication link. In general, a V2V communication concept provides information about the immediately preceding vehicle (e.g., the vehicle directly ahead of the vehicle 500, in the same lane as the vehicle 500), while an I2V communication concept provides information about the traffic further ahead. A CACC system may include either or both I2V and V2V information sources. Given the information of the vehicles ahead of the vehicle 500, CACC may be more reliable and may potentially allow traffic to flow more smoothly and reduce congestion on the roads.

[0135] The FCW system is designed to warn the driver of hazards so that the driver can take corrective action. The FCW system uses a forward-facing camera and / or RADAR sensor 560 coupled to a dedicated processor, DSP, FPGA, and / or ASIC electrically coupled to driver feedback such as a display, speaker, and / or vibration components. The FCW system can provide warnings in the form of audio, visual alerts, vibrations, and / or quick brake pulses.

[0136] An AEB system can detect an imminent forward collision with another vehicle or other object and automatically apply the brakes if the driver does not take corrective action within specified time or distance parameters. The AEB system can use a forward-facing camera and / or RADAR sensor 560 coupled to a dedicated processor, DSP, FPGA, and / or ASIC. When the AEB system detects a hazard, the AEB system typically first alerts the driver to take corrective action to avoid the collision, and if the driver does not take corrective action, the AEB system can automatically apply the brakes as part of an effort to prevent, or at least mitigate, the effects of the predicted collision. The AEB system can include techniques such as Dynamic Brake Support and / or Collision Imminent Braking.

[0137] The LDW system provides visual, audible, and / or tactile warnings, such as vibrations in the steering wheel or seat, to alert the driver when the vehicle 500 crosses a lane marking. The LDW system is not activated when the driver indicates an intentional lane departure by activating a turn signal. The LDW system may use a forward-facing camera coupled to a dedicated processor, DSP, FPGA, and / or ASIC electrically coupled to driver feedback, such as a display, speaker, and / or vibration components.

[0138] The LKA system is a modification of the LDW system, which provides steering input or braking to correct the vehicle 500 if it begins to drift out of its lane.

[0139] The BSW system detects and warns the vehicle driver in the vehicle's blind spot. The BSW system can provide visual, audible, and / or tactile warnings to indicate that merging or changing lanes is unsafe. The system can provide additional warnings when the driver uses a turn signal. The BSW system can use rear-facing cameras and / or RADAR sensors 560 coupled to dedicated processors, DSPs, FPGAs, and / or ASICs electrically coupled to driver feedback, e.g., displays, speakers, and / or vibration components.

[0140] The RCTW system may provide visual, audible, and / or tactile notifications when an object is detected outside the range of the rear camera when the vehicle 500 is backing up. Some RCTW systems include AEB to ensure vehicle brakes are applied to avoid a collision. The RCTW system may use one or more rear-facing RADAR sensors 560 coupled to a dedicated processor, DSP, FPGA, and / or ASIC electrically coupled to driver feedback, e.g., a display, speaker, and / or vibration components.

[0141] Because conventional ADAS systems alert the driver and allow the driver to determine whether a safety condition truly exists and act accordingly, conventional ADAS systems can be prone to producing false positive results that are usually not catastrophic but can be annoying and distracting to the driver. However, in an autonomous vehicle 500, the vehicle 500 itself must decide whether to listen to the results from the primary computer or the secondary computer (e.g., the first controller 536 or the second controller 536) when the results are conflicting. For example, in some embodiments, the ADAS system 538 may be a backup and / or secondary computer to provide perception information to a backup computer rationality module. The backup computer rationality monitor may run redundant and diverse software on hardware components to detect impairments in perception and dynamic driving tasks. Output from the ADAS system 538 may be provided to a supervisory MCU. When the outputs from the primary and secondary computers conflict, the supervisory MCU must decide how to reconcile the conflict to ensure safe operation.

[0142] In some instances, the primary computer may be configured to provide a reliability score to the overseer MCU indicating the reliability of the primary computer in a selected outcome. If the reliability score exceeds a threshold, the overseer MCU may follow the instructions of the primary computer regardless of whether the secondary computers provide conflicting or inconsistent results. If the reliability score does not meet the threshold, and if the primary and secondary computers show different (e.g., conflicting) results, the overseer MCU may arbitrate between the computers to determine the appropriate outcome.

[0143] The supervisory MCU may be configured to execute a neural network that is trained and configured to determine conditions under which the secondary computer provides a false alarm based on the outputs from the primary and secondary computers. Thus, the neural network in the supervisory MCU may learn when the output of the secondary computer may be trusted and when it may not be trusted. For example, when the secondary computer is a RADAR-based FCW system, the neural network in the supervisory MCU may learn when the FCW identifies a metal object that is not actually a hazard, such as a sewer grate or manhole cover, that triggers an alarm. Similarly, when the secondary computer is a camera-based LDW system, the neural network in the supervisory MCU may learn to ignore the LDW when a bicyclist or pedestrian is present and lane departure is, in fact, the safest maneuver. In embodiments that include a neural network executing on the supervisory MCU, the supervisory MCU may include at least one of a DLA or a GPU suitable for executing the neural network with associated memory. In a preferred embodiment, the director MCU may comprise and / or be included as a component of the SoC 504 .

[0144] In other instances, the ADAS system 538 may include a secondary computer that performs ADAS functions using classical rules of computer vision. As such, the secondary computer may use classical computer vision rules (if-then) and the presence of a neural network in the supervisory MCU may improve reliability, safety and performance. For example, the diverse implementations and intentional non-identity may make the overall system more fault-tolerant, especially to failures caused by software (or software-hardware interface) functions. For example, if there is a software bug or error in the software running on the primary computer and non-identical software code running on the secondary computer provides the same overall result, the supervisory MCU may have greater confidence that the overall result is correct and that a bug in the software or hardware on the primary computer has not caused a critical error.

[0145] In some instances, the output of the ADAS system 538 may be fed to the perception block of the primary computer and / or the dynamic driving task block of the primary computer. For example, if the ADAS system 538 indicates a forward collision warning due to an object directly ahead, the perception block may use this information when identifying the object. In other instances, the secondary computer may have its own neural network that is trained as described herein, thus reducing the risk of false positives.

[0146] Vehicle 500 may further include an infotainment SoC 530 (e.g., an in-vehicle infotainment system (IVI)). Although shown and described as an SoC, the infotainment system need not be an SoC and may include two or more separate components. Infotainment SoC 530 may include a combination of hardware and software that may be used to provide audio (e.g., music, personal digital assistant, navigation instructions, news, radio, etc.), video (e.g., TV, movies, streaming, etc.), telephony (e.g., hands-free calling), network connectivity (e.g., LTE, Wi-Fi, etc.), and / or information services (e.g., navigation system, rear park assist, wireless data system, vehicle related information such as fuel level, total distance traveled, brake fuel level, oil level, door open / close, air filter information, etc.) to vehicle 500. For example, the infotainment SoC 530 may be a radio, a disc player, a navigation system, a video player, USB and Bluetooth connectivity, a car computer, in-car entertainment, Wi-Fi, steering wheel audio controls, hands-free voice control, a heads-up display (HUD), an HMI display 534, a telematics device, a control panel (e.g., for controlling and / or interacting with various components, features, and / or systems), and / or other components. The infotainment SoC 530 may further be used to provide information (e.g., visual and / or audible) to a user of the vehicle, such as information from an ADAS system 538, autonomous driving information such as planned vehicle maneuvers, trajectory, surrounding environment information (e.g., intersection information, vehicle information, road information, etc.), and / or other information.

[0147] The infotainment SoC 530 may include GPU functionality. The infotainment SoC 530 may communicate with other devices, systems, and / or components of the vehicle 500 via the bus 502 (e.g., CAN bus, Ethernet, etc.). In some instances, the infotainment SoC 530 may be coupled to the supervisor MCU such that the infotainment system's GPU may perform some self-drive functions in the event that the primary controller 536 (e.g., the vehicle's 500 primary and / or backup computer) fails. In such instances, the infotainment SoC 530 may place the vehicle 500 in a Chauffeur safe shutdown mode, as described herein.

[0148] The vehicle 500 may further include an instrument cluster 532 (e.g., a digital dash, an electronic instrument cluster, a digital instrument panel, etc.). The instrument cluster 532 may include a controller and / or a supercomputer (e.g., a separate controller or a supercomputer). The instrument cluster 532 may include a set of instruments such as a speedometer, a fuel level, an oil pressure, a tachometer, an odometer, a turn signal, a gear shift position indicator, a seat belt warning light, a parking brake warning light, an engine malfunction light, an airbag (SRS) system information, lighting controls, safety system controls, navigation information, etc. In some instances, information may be displayed and / or shared between the infotainment SoC 530 and the instrument cluster 532. In other words, the instrument cluster 532 may be included as part of the infotainment SoC 530, or vice versa.

[0149] 5D is a system diagram of communication between a cloud-based server and an example autonomous vehicle 500 of FIG. 5A according to some embodiments of the disclosure. System 576 may include a server 578, a network 590, and a vehicle including vehicle 500. Server 578 may include multiple GPUs 584(A)-584(H) (collectively referred to herein as GPUs 584), PCIe switches 582(A)-582(H) (collectively referred to herein as PCIe switches 582), and / or CPUs 580(A)-580(B) (collectively referred to herein as CPUs 580). GPUs 584, CPUs 580, and PCIe switches may be interconnected with a high-speed interconnect, such as, but not limited to, an NVLink interface 588 developed by NVIDIA and / or a PCIe connection 586. In some instances, the GPUs 584 are connected via an NVLink and / or NVSwitch SoC, and the GPUs 584 and PCIe switch 582 are connected via a PCIe interconnect. Although eight GPUs 584, two CPUs 580, and two PCIe switches are illustrated, this is not intended to be limiting. Depending on the embodiment, each server 578 may include any number of GPUs 584, CPUs 580, and / or PCIe switches. For example, the servers 578 may each include 8, 16, 32, and / or more GPUs 584.

[0150] Server 578 may receive image data from the vehicle over network 590, the image data representing images showing unexpected or changed road conditions, such as recently started road construction. Server 578 may transmit neural network 592, updated neural network 592, and / or map information 594, including information about traffic and road conditions, to the vehicle over network 590. Updates to map information 594 may include updates to HD map 522, such as information about construction sites, potholes, detours, flooding, and / or other obstacles. In some instances, neural network 592, updated neural network 592, and / or map information 594 may have arisen from new training and / or experience represented in data received from any number of vehicles in the environment and / or based on training performed at a data center (e.g., using server 578 and / or other servers).

[0151] The server 578 may be used to train a machine learning model (e.g., a neural network) based on the training data. The training data may be generated by the vehicle and / or generated in a simulation (e.g., using a game engine). In some instances, the training data is tagged (e.g., if the neural network benefits from supervised learning) and / or undergoes other pre-processing, while in other instances, the training data is not tagged and / or pre-processed (e.g., if the neural network does not require supervised learning). The training may be performed according to any one or more classes of machine learning techniques, including, but not limited to, the following classes: supervised training, semi-supervised training, unsupervised training, self-learning, reinforcement learning, federated learning, transfer learning, feature learning (including principal component and cluster analysis), multi-linear subspace learning, manifold learning, representation learning (including preliminary dictionary learning), rule-based machine learning, anomaly detection, and variations or combinations thereof. After the machine-learned model has been traced, the machine-learned model may be used by the vehicle (e.g., transmitted to the vehicle via network 590) and / or the machine-learned model may be used by server 578 to remotely monitor the vehicle.

[0152] In some instances, server 578 can receive data from vehicles and apply the data to state-of-the-art real-time neural networks for real-time intelligent inference. Server 578 can include deep learning supercomputers and / or dedicated AI computers powered by GPU 584, such as DGX and DGX Station machines developed by NVIDIA. However, in some instances, server 578 can include a deep learning infrastructure that uses only CPU-powered data centers.

[0153] The deep learning infrastructure of server 578 may be capable of rapid real-time inference and may use that capability to evaluate and validate the health of the processor, software, and / or associated hardware within vehicle 500. For example, the deep learning infrastructure may receive periodic updates from vehicle 500, such as a sequence of images and / or objects where vehicle 500 was located within the sequence of images (e.g., via computer vision and / or other machine learning object classification techniques). The deep learning infrastructure may run its own neural networks to identify objects and compare them to objects identified by vehicle 500, and if the results do not match and the infrastructure concludes that the AI ​​within vehicle 500 is not functioning properly, server 578 may send a signal to vehicle 500 that infers control, notifies passengers, and commands the failsafe computer of vehicle 500 to complete a safe parking maneuver.

[0154] For inference, the server 578 may include a GPU 584 and one or more programmable inference accelerators (e.g., NVIDIA's TensorRT). The combination of a GPU-powered server and inference acceleration can enable real-time responsiveness. In other instances, such as when less performance is required, servers powered by CPUs, FPGAs, and other processors may be used for inference.

[0155] Exemplary Computing Device 6 is a block diagram of an example computing device 600 suitable for use in implementing some embodiments of the present disclosure. The computing device 600 may include an interconnect system 602 that indirectly or directly couples the following devices: memory 604, one or more central processing units (CPUs) 606, one or more graphic processing units (GPUs) 608, a communication interface 610, input / output (I / O) ports 612, input / output components 614, a power supply 616, one or more presentation components 618 (e.g., displays), and one or more logic units 620. In at least one embodiment, the computing device 600 may include one or more virtual machines (VMs) and / or any of its components may include virtual components (e.g., virtual hardware components). As non-limiting examples, one or more of GPUs 608 may include one or more vGPUs, one or more of CPUs 606 may include one or more vCPUs, and / or one or more of logic units 620 may include one or more virtual logic units. As such, computing device 600 may include discrete components (e.g., an entire GPU dedicated to computing device 600), virtual components (e.g., a portion of a GPU dedicated to computing device 600), or a combination thereof.

[0156] Although the various blocks in FIG. 6 are shown connected via interconnect system 602 with lines, this is not intended to be limiting and is merely for clarity. For example, in some embodiments, a presentation component 618, such as a display device, may be considered an I / O component 614 (e.g., where the display is a touch screen). As another example, CPU 606 and / or GPU 608 may include memory (e.g., memory 604 may represent a storage device in addition to the memory of GPU 608, CPU 606, and / or other components). In other words, the computing devices of FIG. 6 are merely exemplary. Categories such as "workstations," "servers," "laptops," "desktops," "tablets," "client devices," "mobile devices," "handheld devices," "gaming consoles," "electronic control units (ECUs)," "virtual reality systems," and / or other device or system types are all intended to be within the scope of the computing devices of FIG. 6 and therefore are not to be differentiated.

[0157] The interconnect system 602 may represent one or more links or buses, such as an address bus, a data bus, a control bus, or a combination thereof. The interconnect system 602 may include one or more bus or link types, such as an industry standard architecture (ISA) bus, an extended industry standard architecture (EISA) bus, a video electronics standards association (VESA) bus, a peripheral component interconnect (PCI) bus, a peripheral component interconnect express (PCIe) bus, and / or another type of bus or link. In some embodiments, there are direct connections between components. As an example, the CPU 606 may be directly connected to the memory 604. Additionally, the CPU 606 may be directly connected to the GPU 608. When there are direct or point-to-point connections between components, the interconnect system 602 may include a PCIe link to implement the connections. In these examples, a PCI bus need not be included in the computing device 600.

[0158] Memory 604 may include any of a variety of computer readable media. Computer readable media may be any available media that can be accessed by computing device 600. Computer readable media may include both volatile and nonvolatile media, and removable and non-removable media. By way of example, and not limitation, computer readable media may comprise computer storage media and communication media.

[0159] Computer storage media may include both volatile and nonvolatile media and / or removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules, and / or other data types. For example, memory 604 may store computer readable instructions (e.g., representing programs and / or program elements), such as an operating system. Computer storage media may include, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium that may be used to store the desired information and that may be accessed by computing device 600. As used herein, computer storage media does not include the signals themselves.

[0160] Computer storage media may embody computer-readable instructions, data structures, program modules, and / or other data types in a modulated data signal such as a carrier wave or other transport mechanism and include any information delivery media. The term "modulated data signal" may refer to a signal that has one or more of its characteristics set or changed in such a manner as to encode information in the signal. By way of example, and not limitation, computer storage media may include wired media, such as a wired network or direct-wired connection, and wireless media, such as acoustic, RF, infrared and other wireless media. Combinations of any of the above should also be included within the scope of computer-readable media.

[0161] The CPU 606 may be configured to execute at least some of the computer readable instructions to control one or more components of the computing device 600 to perform one or more of the methods and / or processes described herein. The CPU 606 may include one or more (e.g., 1, 2, 4, 8, 28, 72, etc.) cores each capable of processing multiple software threads simultaneously. The CPU 606 may include any type of processor, and may include different types of processors depending on the type of computing device 600 implemented (e.g., a processor with fewer cores for a mobile device and a processor with more cores for a server). For example, depending on the type of computing device 600, the processor may be an Advanced RISC Machines (ARM) processor implemented using Reduced Instruction Set Computing (RISC) or an x86 processor implemented using Complex Instruction Set Computing (CISC). Computing device 600 may include one or more CPUs 606 within one or more microprocessors or auxiliary coprocessors, such as computational coprocessors.

[0162] In addition to or instead of CPU 606, GPU 608 may be configured to execute at least some of the computer readable instructions to control one or more components of computing device 600 to perform one or more of the methods and / or processes described herein. One or more of GPUs 608 may be integrated GPUs (e.g., with one or more of CPUs 606 and / or one or more of GPUs 608 may be discrete GPUs. In an embodiment, one or more of GPUs 608 may be coprocessors of one or more of CPUs 606. GPU 608 may be used by computing device 600 to render graphics (e.g., 3D graphics) or perform general purpose computations. For example, GPU 608 may be configured to perform general purpose computing on a GPU (GPGPU). A graphics processing unit (GPU) may be used for processing graphics data. GPU 608 may include hundreds or thousands of cores capable of processing hundreds or thousands of software threads simultaneously. GPU 608 may generate pixel data for an output image in response to rendering commands (e.g., rendering commands from CPU 606 received via a host interface). GPU 608 may include graphics memory, e.g., display memory, for storing the pixel data or any other suitable data, e.g., GPGPU data. The display memory may be included as part of memory 604. GPU 608 may include two or more GPUs operating in parallel (e.g., via links). The links may connect the GPUs directly (e.g., using NVLINK) or may connect the GPUs via a switch (e.g., using NVSwitch). When coupled together, each GPU 608 may generate pixel data or GPGPU data for a different portion of the output or for a different output (e.g., a first GPU for a first image and a second GPU for a second image). Each GPU may include its own memory or may share memory with the other GPUs.

[0163] In addition to or instead of the CPU 606 and / or GPU 608, the logic unit 620 may be configured to execute at least some of the computer readable instructions to control one or more of the computing devices 600 to perform one or more of the methods and / or processes described herein. In an embodiment, the CPU 606, the GPU 608, and / or the logic unit 620 may execute any combination of the methods, processes and / or portions thereof, either discretely or jointly. One or more of the logic units 620 may be part of and / or integrated with one or more of the CPU 606 and / or GPU 608, and / or one or more of the logic units 620 may be a discrete component to or otherwise external to the CPU 606 and / or GPU 608. In an embodiment, one or more of the logic units 620 may be a co-processor of one or more of the CPU 606 and / or GPU 608.

[0164] Examples of logic unit 620 include one or more processing cores and / or components thereof, such as a Tensor Core (TC), a Tensor Processing Unit (TPU), a Pixel Visual Core (PVC), a Vector Processing Unit (VPU), a Graphics Processing Cluster (GPC), a Texture Processing Cluster (TPC), a Streaming Multiprocessor (SM), a Tree Traversal Unit (TTU), an Artificial Intelligence Accelerator (AIA), a Deep Learning Accelerator (DLA), an Arithmetic Logic Unit (ALU), an Application Specific Integrated Circuit (ASIC), a Floating Point Unit (FPU), an Input / Output (I / O) element, a Data Processing Unit (DPU), a 3D Array (Array ... The term "PCI" refers to a set of peripheral component interconnect (PCI) or peripheral component interconnect express (PCIe) elements, and / or the like.

[0165] The communications interface 610 may include one or more receivers, transmitters, and / or transceivers that enable the computing device 600 to communicate with other computing devices over electronic communications networks, including wired and / or wireless communications. The communications interface 610 may include components and functionality to enable communication over any of a number of different networks, such as a wireless network (e.g., Wi-Fi, Z-Wave, Bluetooth, Bluetooth LE, ZigBee, etc.), a wired network (e.g., communicating over Ethernet or InfiniBand), a low power wide area network (e.g., LoRaWAN, SigFox, etc.), and / or the Internet.

[0166] The I / O ports 612 may enable the computing device 600 to be logically coupled to other devices, including I / O components 614, presentation components 618, and / or other components, some of which may be built-in (e.g., integrated) to the computing device 600. Exemplary I / O components 614 include a microphone, a mouse, a keyboard, a joystick, a game pad, a game controller, a satellite dish, a scanner, a printer, a wireless device, and the like. The I / O components 614 may provide a natural user interface (NUI) that processes air gestures, voice, or other physiological input generated by a user. In some cases, the input may be sent to an appropriate network element for further processing. The NUI may implement any combination of voice recognition, stylus recognition, facial recognition, biometric recognition, gesture recognition on and next to the screen, air gestures, head and eye tracking, and touch recognition in conjunction with the display of the computing device 600 (as described in more detail below). Computing device 600 may include a depth camera, such as a stereoscopic camera system, an infrared camera system, an RGB camera system, touch screen technology, and combinations thereof, for gesture detection and recognition. Additionally, computing device 600 may include an accelerometer or gyroscope (e.g., as part of an inertia measurement unit (IMU)) to enable detection of motion. In some instances, the output of the accelerometer or gyroscope may be used by computing device 600 to render immersive augmented or virtual reality.

[0167] The power supply 616 may include a hardwired power supply, a battery power supply, or a combination thereof. The power supply 616 may provide power to the computing device 600 to enable the components of the computing device 600 to operate.

[0168] The presentation component 618 may include a display (e.g., a monitor, a touch screen, a television screen, a heads-up display (HUD), other display types, or a combination thereof), speakers, and / or other presentation components. The presentation component 618 can receive data from other components (e.g., GPU 608, CPU 606, etc.) and output data (e.g., as images, video, sound, etc.).

[0169] Exemplary Data Center 7 illustrates an example data center 700 that may be used in at least one embodiment of the present disclosure. The data center 700 may include a data center infrastructure layer 710, a framework layer 720, a software layer 730, and / or an application layer 740.

[0170] 7, the data center infrastructure layer 710 may include a resource orchestrator 712, grouped computational resources 714, and node computational resources ("node CRs") 716(1)-716(N), where "N" represents any integer, natural number. In at least one embodiment, the node CRs 716(1)-716(N) may include, but are not limited to, any number of central processing units ("CPUs") or other processors (including accelerators, field programmable gate arrays (FPGAs), graphics processors or graphics processing units (GPUs), etc.), memory devices (e.g., dynamic read-only memory), storage devices (e.g., solid state or disk drives), network input / output ("NW I / O") devices, network switches, virtual machines ("VMs"), power modules, and / or cooling modules, etc. In some embodiments, one or more of the nodes CR716(1)-716(N) may correspond to a server having one or more of the aforementioned computing resources. In addition, in some embodiments, the nodes CR716(1)-716(N) may include one or more virtual components, such as a vGPU, a vCPU, and / or the like, and / or one or more of the nodes CR716(1)-716(N) may correspond to a virtual machine (VM).

[0171] In at least one embodiment, the grouped computing resources 714 may include separate groups of nodes CR716 housed in one or more racks (not shown), or multiple racks housed in data centers in various geographic locations (also not shown). The separate groups of nodes CR716 in the grouped computing resources 714 may include grouped computing, network, memory, or storage resources that may be configured or assigned to support one or more workloads. In at least one embodiment, several nodes CR716 including CPUs, GPUs, and / or other processors may be grouped in one or more racks to provide computing resources to support one or more workloads. The one or more racks may also include any number of power modules, cooling modules, and / or network switches in any combination.

[0172] The resource orchestrator 722 may configure or otherwise control one or more nodes CR 716(1)-716(N) and / or grouped computational resources 714. In at least one embodiment, the resource orchestrator 722 may include a software design infrastructure ("SDI") management entity of the data center 700. The resource orchestrator 722 may include hardware, software, or some combination thereof.

[0173] In at least one embodiment, as shown in FIG. 7 , framework layer 720 may include a job scheduler 732, a configuration manager 734, a resource manager 736, and / or a distributed file system 738. Framework layer 720 may include a framework to support software 732 in software layer 730 and / or one or more applications 742 in application layer 740. Software 732 or applications 742 may include web-based service software or applications, such as those offered by Amazon Web Services, Google Cloud, and Microsoft Azure, respectively. Framework layer 720 may be a type of free and open source software web application framework, such as, but not limited to, Apache Spark™ (hereinafter “Spark”), that may use distributed file system 738 for large-scale data processing (e.g., “big data”). In at least one embodiment, job scheduler 732 may include a Spark driver to facilitate scheduling of workloads supported by various tiers of data center 700. The configuration manager 734 may be capable of configuring different layers, for example, the software layer 730 and the framework layer 720 including Spark and distributed file system 738 to support large scale data processing. The resource manager 736 may be capable of managing the clustered or grouped computing resources that are mapped or allocated to support the distributed file system 738 and the job scheduler 732. In at least one embodiment, the clustered or grouped computing resources may include the computing resources 714 grouped in the data center infrastructure layer 710. The resource manager 1036 may coordinate with the resource orchestrator 712 to manage these mapped or allocated computing resources.

[0174] In at least one embodiment, software 732 included in software layer 730 may include software used by at least a portion of nodes CR 716(1)-716(N), grouped computational resources 714, and / or distributed file system 738 of framework layer 720. The one or more types of software may include, but are not limited to, Internet web page searching software, email virus scanning software, database software, and streaming video content software.

[0175] In at least one embodiment, the applications 742 included in the application layer 740 may include one or more types of applications used by at least a portion of the nodes CRs 716(1)-716(N), the grouped computational resources 714, and / or the distributed file system 738 of the framework layer 720. The one or more types of applications may include, but are not limited to, any number of genomics applications, cognitive computing, and machine learning applications, including training or inference software, machine learning framework software (e.g., PyTorch, TensorFlow, Caffe, etc.), and / or other machine learning applications used in conjunction with one or more embodiments.

[0176] In at least one embodiment, any of configuration manager 734, resource manager 736, and resource orchestrator 712 can implement any number and type of self-rewriting actions based on any amount and type of data obtained in any technically possible manner. The self-rewriting actions can free data center operators of data center 700 from making potentially poor configuration decisions and possibly avoiding underutilized and / or underperforming portions of the data center.

[0177] Data center 700 may include tools, services, software, or other resources to train one or more machine learning models or to predict or infer information using one or more machine learning models according to one or more embodiments described herein. For example, a machine learning model may be trained by calculation of weight parameters via a neural network architecture using the software and / or computing resources described above with respect to data center 700. In at least one embodiment, a trained or deployed machine learning model corresponding to one or more neural networks may be used to infer or predict information using the resources described above with respect to data center 700, for example, by using weight parameters calculated via one or more training techniques, including but not limited to those described herein.

[0178] In at least one embodiment, data center 700 may use CPUs, application specific integrated circuits (ASICs), GPUs, DPUs, FPGAs, and / or other hardware (or corresponding virtual computing resources) for training and / or performing inference using the aforementioned resources. Additionally, one or more of the aforementioned software and / or hardware resources may be configured as services, such as image recognition, speech recognition, or other artificial intelligence services, to enable a user to train or perform inference on information.

[0179] Example Network Environment A network environment suitable for use in implementing embodiments of the present disclosure may include one or more client devices, servers, network attached storage (NAS), other back-end devices, and / or other device types. The client devices, servers, and / or other device types (e.g., each device) may be implemented with one or more instances of computing device 600 of FIG. 6, e.g., each device may include similar components, features, and / or functionality of computing device 600. In addition, if a back-end device (e.g., server, NAS, etc.) is implemented, the back-end device may be included as part of a data center 700, examples of which are further detailed herein with respect to FIG.

[0180] Components of a network environment may communicate with each other via a network, which may be wired, wireless, or both. A network may include multiple networks or a network of networks. Illustratively, a network may include one or more wide area networks (WANs), one or more local area networks (LANs), one or more public networks, such as the Internet and / or the Public Switched Telephone Network (PSTN), and / or one or more private networks. When a network includes a wireless telecommunications network, components such as base stations, communication towers, or access points (as well as other components) may provide wireless connectivity.

[0181] Compatible network environments may include one or more peer-to-peer network environments (wherein a server may not be included in the network environment) and one or more client-server network environments (wherein a server or servers may be included in the network environment). In a peer-to-peer network environment, the functionality described herein with respect to a server may be implemented in any number of client devices.

[0182] In at least one embodiment, the network environment may include one or more cloud-based network environments, distributed computing environments, combinations thereof, and the like. The cloud-based network environment may include a framework layer, a job scheduler, a resource manager, and a distributed file system implemented on one or more of the servers, which may include one or more core network servers and / or edge servers. The framework layer may include a framework to support software in the software layer and / or one or more applications in the application layer. The software or applications may include web-based service software or applications, respectively. In an embodiment, one or more of the client devices may use the web-based service software or applications (e.g., by accessing the service software and / or applications via one or more application programming interfaces (APIs)). The framework layer may be, but is not limited to, a type of free and open source software web application framework that may use a distributed file system for large-scale data processing (e.g., "big data"), for example.

[0183] A cloud-based network environment may provide cloud computing and / or cloud storage performing any combination of the computing and / or data storage functions (or one or more portions thereof) described herein. Any of these various functions may be distributed across multiple locations from a central or core server (e.g., one or more data centers that may be distributed across a state, territory, country, or world). When a connection to a user (e.g., a client device) is relatively close to an edge server, the core server may delegate at least a portion of the functionality to the edge server. A cloud-based network environment may be private (e.g., restricted to a single organization), public (e.g., available to multiple organizations), and / or a combination thereof (e.g., a hybrid cloud environment).

[0184] A client device may include at least some of the components, features, and functionality of the exemplary computing device 600 described herein with respect to Figure 6. By way of illustration, and not by way of limitation, a client device may be embodied as a personal computer (PC), a laptop computer, a mobile device, a smartphone, a tablet computer, a smart watch, a wearable computer, a personal digital assistant (PDA), an MP3 player, a virtual reality headset, a global positioning system (GPS) or device, a video player, a video camera, a surveillance device or system, a vehicle, a boat, an airship, a virtual machine, a drone, a robot, a handheld communications device, a hospital device, a gaming device or system, an entertainment system, a vehicle computer system, an embedded system controller, a remote control, an appliance, a consumer electronic device, a workstation, an edge device, any combination of these depicted devices, or any other suitable device.

[0185] The present disclosure may be described in the general context of computer code or machine usable instructions, including computer executable instructions, such as program modules, being executed by a computer or other machine, such as a personal digital assistant or other handheld device. Generally, program modules, including routines, programs, objects, components, data structures, etc., refer to code that performs particular tasks or implements particular abstract data types. The present disclosure may be implemented in a variety of configurations, including handheld devices, consumer electronics, general purpose computers, more specialized computing devices, etc. The present disclosure may also be implemented in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network.

[0186] As used herein, the statement "and / or" with respect to two or more elements should be interpreted to mean only one element or a combination of elements. For example, "element A, element B, and / or element C" may include element A only, element B only, element C only, elements A and element B, elements A and element C, elements B and element C, or elements A, B, and C. In addition, "at least one of element A or element B" may include at least one of element A, at least one of element B, or at least one of element A and at least one of element B. Furthermore, "at least one of element A and element B" may include at least one of element A, at least one of element B, or at least one of element A and at least one of element B.

[0187] The subject matter of the present disclosure has been described with specificity to meet statutory requirements. However, the description itself is not intended to limit the scope of the present disclosure. Instead, the inventors intend that the claimed subject matter may be implemented in other ways, including different steps or combinations of steps similar to those described in this document, in conjunction with other current or future technologies. Furthermore, although the terms "step" and / or "block" may be used herein to connote different elements of the method used, these terms should not be construed as implying any particular order between the various steps disclosed herein unless and when the order of the individual steps is explicitly described.

Claims

1. applying the first data to a deep neural network (DNN) and a patch neural network (PNN); processing the first data to calculate second data, The processing step includes: computing a first output based at least in part on the first data using a first subset of nodes of the DNN; processing the first output using a second subset of the nodes of the DNN; computing a second output based at least in part on the first data using a first subset of nodes of the PNN; processing the first output and the second output using a second subset of the nodes of the PNN; and performing one or more actions based at least in part on the second data. Including, processing the second data, the second data representing a final output of the DNN and a final output of the PNN; determining an output selection based at least in part on analyzing the final output of the DNN and the final output of the PNN; A method comprising: The method, wherein the step of determining the output selection is performed using a gating neural network (GNN) that outputs a binary decision indicating whether to use the final output of the DNN or the final output of the PNN as the output selection.

2. The method of claim 1 , wherein a PNN layer of the PNN that contains the nodes of the PNN is separate from a DNN layer of the DNN that contains the nodes of the DNN.

3. The method of claim 1 , wherein one or more layers of a larger DNN comprise the nodes of the PNN and the nodes of the DNN.

4. The method of claim 1 , wherein the final output of the PNN includes at least one output type that is represented by the final output of the DNN and at least one output type that is not represented by the final output of the DNN.

5. The method of claim 1, wherein the step of performing one or more actions is based at least in part on the output selection.

6. the analyzing step includes analyzing a first confidence value associated with the final output of the DNN and a second confidence value associated with the final output of the PNN; The method of claim 5 , wherein determining the output selection comprises selecting the output selection based at least in part on analyzing the first reliability value and the second reliability value.

7. The method of claim 1 , further comprising determining to use the final output of the DNN or the final output of the PNN based at least in part on one or more received signals.

8. The method of claim 7 , wherein the one or more received signals correspond to a location, weather conditions, lighting conditions, or time of day.

9. The method further comprising: A control system for an autonomous or semi-autonomous machine; Perception systems for autonomous or semi-autonomous machines; A system for performing a simulation operation; A system for performing deep learning operations, A system implemented using edge devices; A system implemented using a robot, A system incorporating one or more virtual machines (VMs); A system implemented at least in part in a data center; or System implemented at least in part using cloud computing resources - Patents.com The method of claim 1 , wherein the method is performed by at least one of:

10. The method of claim 1 , wherein a first parameter of the node of the PNN is trained while a second parameter of the node of the DNN is fixed.

11. The method of claim 1 , wherein weights associated with relationships between the second output and one or more of the nodes of the DNN are fixed to zero.

12. 1. A system comprising: a system on a chip (SoC) for generating an output for controlling an autonomous vehicle, the SoC comprising: A processor for executing a neural network wherein the neural network comprises: a first layer including a first set of initial nodes for processing first sensor data to calculate a first output and a first set of patch nodes for processing the first sensor data to calculate a second output; a second layer following the first layer, the second layer including a second set of initial nodes for processing the first output to compute a third output, and a second set of patch nodes for processing the first output and the second output to compute a fourth output; Including, an output for controlling the autonomous vehicle is selected using a gating neural network (GNN) that outputs a binary decision based at least in part on the third output and the fourth output.

13. 13. The system of claim 12, wherein the first set of nodes is trained separately from the second set of nodes, and parameters associated with the first set of nodes are fixed during training of the second set of nodes.

14. 13. The system of claim 12, wherein one or more zeros are added to a matrix corresponding to the second set of initial nodes to match a dimensionality of a matrix corresponding to the second set of patch nodes, the one or more zeros corresponding to a weight between a relationship of the second output of the first set of patch nodes and the second set of initial nodes.

15. the neural network is trained with the first set of initial nodes as a first separate layer and the first set of patch nodes as a second separate layer; 13. The system of claim 12, wherein after training, the first layer is generated by concatenating weights associated with the first set of initial nodes and weights associated with the first set of patch nodes.

16. 13. The system of claim 12, wherein during training, the neural network is subjected to a pruning operation to remove one or more sets of patch nodes corresponding to one or more layers of the neural network.

17. applying the first data to a deep neural network (DNN) and a patch neural network (PNN); Processing the first data to calculate second data 11. A processor comprising one or more circuits for: computing a first output based at least in part on the first data using a first subset of nodes of the DNN; processing the first output using a second subset of the nodes of the DNN; computing a second output based at least in part on the first data using a first subset of nodes of the PNN; processing the first output and the second output using a second subset of the nodes of the PNN; and performing one or more actions based at least in part on the second data; the second data represents a final output of the DNN and a final output of the PNN; an output selection is determined based at least in part on an analysis of the final output of the DNN and the final output of the PNN; A processor, wherein the output selection is determined using a Gating Neural Network (GNN) that outputs a binary decision indicating whether to use the final output of the DNN or the final output of the PNN.

18. 20. The processor of claim 17, wherein the nodes of the PNN correspond to one or more layers of the PNN that are separate from one or more layers of the DNN.

19. 20. The processor of claim 17, wherein the nodes of the PNN are included in addition to the nodes of the DNN to generate a combined layer.

20. The processor, A control system for an autonomous or semi-autonomous machine; Perception systems for autonomous or semi-autonomous machines; A system for performing a simulation operation; A system for performing deep learning operations, A system implemented using edge devices; A system implemented using a robot, A system incorporating one or more virtual machines (VMs); A system implemented at least in part in a data center; or A system implemented at least in part using cloud computing resources; The processor of claim 17 .

21. 20. The processor of claim 17, wherein a first parameter of the node of the PNN is trained while a second parameter of the node of the DNN is fixed.

22. 20. The processor of claim 17, wherein a weight associated with a relationship between the second output of the node of the PNN and the node of the DNN is fixed to zero.

23. 20. The processor of claim 17, wherein the PNN is verified at least in part by running the PNN in a shadow mode.

Images