OPTIMIZATION AND UPDATE SYSTEM FOR DEEP LEARNING MODELS

Abstract

Traditionally, a software application is developed, tested, and then published for use to end users. Any subsequent update made to the software application is generally in the form of a human programmed modification made to the code in the software application itself, and further only becomes usable once tested and published by developers and/or publishers, and installed by end users having the previous version of the software application. This typical software application lifecycle causes delays in not only generating improvements to software applications, but also to those improvements being made accessible to end users. To help avoid these delays and improve performance of software applications, deep learning models may be made accessible to the software applications for use in performing inferencing operations to generate inferenced data output for the software applications, which the software applications may then use as desired. These deep learning models can furthermore be improved independently of the software applications using manual and/or automated processes.

Description

CLAIM OF PRIORITY

This application claims the benefit of U.S. Provisional Application No. 62/717,735, titled “CONTINUOUS OPTIMIZATION AND UPDATE SYSTEM FOR DEEP LEARNING MODELS,” filed Aug. 10, 2018, the entire contents of which is incorporated herein by reference.

RELATED APPLICATIONS

This application is related to co-pending U.S. application Ser. No. ______, titled “DEEP LEARNING MODEL EXECUTION USING TAGGED DATA” (Attorney Ref: NVIDP1276/18-SC-0202US01) filed Aug. ______, 2019, the entire contents of which is incorporated herein by reference.

This application is related to co-pending U.S. application Ser. No. ______, titled “AUTOMATIC DATASET CREATION USING SOFTWARE TAGS” (Attorney Ref: NVIDP1277/18-SC-0197US01) and filed Aug. ______, 2019, the entire contents of which is incorporated herein by reference.

TECHNICAL FIELD

The present disclosure relates to deep learning used by software applications.

BACKGROUND

Traditionally, a software application is developed, tested, and then published for use to end users. Any subsequent update made to the software application is generally in the form of a human programmed modification made to the code in the software application itself, and further only becomes usable once tested and published by an application developer and publisher, and installed by end users having the previous version of the software application. This typical software application lifecycle causes delays in not only generating improvements to software applications, but also to those improvements being made accessible to end users.

There is a need for addressing these issues and/or other issues associated with the prior art.

SUMMARY

A method, computer readable medium, and system are disclosed for improving deep learning models that perform inferencing operations to provide inferenced data to software applications. In an embodiment, a deep learning model usable for performing inferencing operations and for providing inferenced data is stored. Additionally, the deep learning model is updated to create an updated version of the deep learning model. Further, the updated version of the deep learning model is distributed to a client for use in providing the inferenced data.

In another embodiment, a deep learning model is stored. Additionally, the deep learning model is executed to perform inferencing operations and to provide inferenced data to a software application. Further, an updated version of the deep learning model is received. Still yet, the updated version of the deep learning model is executed to provide additional inferenced data to the software application.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 illustrates a block diagram of a system including a server that provisions a deep learning model to a client for use by a software application installed on the client, in accordance with an embodiment.

FIG. 2 illustrates a flowchart of a server method for improving a deep learning model for use by a client, in accordance with an embodiment.

FIG. 3 illustrates a flowchart of a client method for implementing an improved deep learning model that provides inferenced data to a local software application, in accordance with an embodiment.

FIG. 4A illustrates a block diagram of a system 400 for updating a deep learning model that performs inferencing operations and provides inferenced data to a software application, in accordance with an embodiment.

FIG. 4B illustrates a flowchart of the method of the client of FIG. 4, in accordance with an embodiment.

FIG. 5A illustrates inference and/or training logic, according to at least one embodiment;

FIG. 5B illustrates inference and/or training logic, according to at least one embodiment;

FIG. 6 illustrates training and deployment of a neural network, according to at least one embodiment;

FIG. 7 illustrates an example data center system, according to at least one embodiment;

DETAILED DESCRIPTION

FIG. 1 illustrates a block diagram of a system 100 including a server 101 that provisions a deep learning model 102 to a client 103 for use by a software application 104 installed on the client 103, in accordance with an embodiment.

With respect to the present description, the server 101 may be any computing device, virtualized computing device, or combination of devices, capable of communicating with the client 103 over a wired or wireless connection, for the purpose of provisioning the deep learning model 102 to the client 103 for use by a software application 104 installed on the client 103. For example, the server 101 may include a hardware memory (e.g. random access memory (RAM), etc.) for storing the deep learning model 102 and a hardware processor (e.g. central processing unit (CPU), graphics processing unit (GPU), etc.) for provisioning the deep learning model 102 from the memory to the client 103 over the wired or wireless connection. The server 101 may provision the deep learning model 102 to the client 103 by sending a copy of the deep learning model 102 over the wired or wireless connection to the client 103.

Also with respect to the present description, the client 103 may be any computing device (including, without limitation, computing devices that are wholly or partially virtualized) capable of communicating with the server 101 over the wired or wireless connection, for the purpose of receiving from the server 101 the deep learning model 102 for use by the software application 104 installed on the client 103. Thus, the client 103 may not necessarily be an end-user device (e.g. personal computer, laptop, mobile phone, etc.) but may also be a server or other cloud-based computer system having the software application 104 installed thereon. In the case where the client 103 is a cloud-based computer system, output of the software application 104 may optionally be streamed or otherwise communicated to an end-user device. Generally, the client 103 may include a memory for storing the deep learning model 102 and a processor by which the software application 104 installed on the client 103 uses the deep learning model 102 for obtaining inferenced data. By storing a copy of the deep learning model 102 at the client (e.g. on a hard drive of the client), the client executes the deep learning model 102 locally.

The deep learning model 102 is a machine learned network (e.g. deep neural network) that is trained to perform inferencing operations and to provide inferenced data from input data. The deep learning model 102 may be trained using supervised or unsupervised training techniques. Optionally, the server 101 may be used to perform the training of the deep learning model 102, or may receive the already trained deep learning model 102 from another device.

The deep learning model 102 may be trained for performing any desired type of inferencing and making any desired type of inferences. However, in the present embodiment, the deep learning model 102 outputs inferences that are usable by the software application 104 installed on the client 103. It should be noted that the deep learning model 102 may similarly be used by other software applications which may be installed on the client 103 or other clients, and thus may not necessarily be specifically trained for use by the software application 104 but instead may be trained more generically for use by multiple different software applications. In any case, the deep learning model 102 may not be coded within the software application 104 itself, but may be accessible to the software application 104 as external functionality (e.g. as a software patch) via an application programming interface (API). As a result, the deep learning model 102 may not necessarily be developed and provided by a same developer of the software application 104 but instead may be developed and provided by a third-party developer.

In the present embodiment, the software application 104 installed on the client 103 provides input data to the deep learning model 102 which processes the input data to perform inferencing and/or to return one or more inferences (i.e. inferenced data) for the input data. Accordingly, the deep learning model 102 is trained to process the input data and make inferences therefrom. The inferenced data is output by the deep learning model 102 to the software application 104 for use by functions, tasks, etc. of the software application 104.

There are various use cases for the system 100 described above. In one embodiment, the software application 104 may be a video game, virtual reality application, image classification and other processing, sensor data analysis, or other graphics-related computer program. In this embodiment, the deep learning model 102 may provide certain image-related inferences, such as providing from an input image or other input data an anti-aliased image, an image with upscaled resolution, a denoised image, and/or any other output image that is modified in at least one respect from the input image or other input data. As another example, the deep learning model 102 may provide inference output that can be used to apply certain video-related effects, such as providing from input video or other input data a slow-motion version of the input video or other input data, a super sampling of the input video or other input data, etc.

In another embodiment, the software application 104 may be a voice recognition application or other audio-related computer program. In this embodiment, the deep learning model 102 may provide inference output that can be used to apply certain audio-related effects, such as providing from an input audio or other input data a language translation, a voice recognized command, and/or any other output that is inferenced from the input audio or other input data.

The system 100 configuration described above enables improvements to be made to the deep learning model 102 without necessarily requiring any changes within the software application 104 itself. Thus, the software application 104 may inherently benefit from the improvements made to the deep learning model 102, and thus an end-user or other system using the software application 104 may benefit from the improvements made to the deep learning model 102, without the tradeoff of the usual delays associated with updating the software application 104 itself. All that may be required is that the copy of the deep learning model 102 on the client 103 be updated to the improved version.

For example, when the deep learning model 102 is improved to be faster, to be less computation-intensive, and/or to provide more accurate inferences, the software application 104 may inherently be improved by way of its use of the deep learning model 102 during execution thereof. For example, the software application 104 may likewise provide faster results, results with less computations, and/or more accurate results as a result of its use of the improved deep learning model 102.

The embodiments below describe systems and methods specifically for improving deep learning models that provide inferenced data to software applications. It should be noted that the systems and methods described below may be implemented in the context of the system 100 of FIG. 1.

FIG. 2 illustrates a flowchart of a server method 200 for improving a deep learning model for use by a client, in accordance with an embodiment. Accordingly, in one embodiment, the method 200 may be performed by the server 101 of FIG. 1.

In operation 201, a deep learning model is stored. In the context of the present method 200, the deep learning model is usable for performing inferencing operations and/or providing inferenced data to a software application (e.g. such as the deep learning model 102 used by the software application 104 of FIG. 1). The deep learning model may be stored locally (e.g. by the server 101). In one embodiment, the deep learning model may be stored in a local repository with other deep learning models usable for performing inferencing operations and/or providing other types of inferenced data to the software application or other software applications.

In operation 202, the deep learning model is updated to create an improved (updated) version of the deep learning model. It should be noted that any aspect(s) of the deep learning model may be updated to create the improved version of the deep learning model. In any case though, the update to the deep learning model improves (e.g. optimizes) the deep learning model in at least one respect.

In particular, the deep learning model may be updated by retraining the deep learning model and/or reconfiguring the deep learning model with new parameters (e.g., weights) or hyperparameters. The updating may be performed automatically by software and/or other neural networks. Thus, the process of updating the deep learning model to create an improved version thereof may be performed without requiring user intervention.

In one embodiment, as noted above, the deep learning model may be retrained, specifically using a changed dataset. For example, where the deep learning model was last trained using a particular dataset, the deep learning model may be retrained using a dataset that is changed from the particular dataset. The changed dataset may include additional data that was not included in the particular dataset that was last used to train the deep learning model and/or may remove data that was included in the particular dataset.

In another embodiment, as noted above, the deep learning model may be updated with one or more reconfigurations being made to the deep learning model. With respect to the option to reconfigure the deep learning model, the deep learning model may be updated according to a hyperparameter adjustment. In the context of the present description, a hyperparameter refers to a parameter whose value is used to control the learning process for the deep learning model (as opposed to the values of other parameters that are learned). For example, where the deep learning model was last trained according to a particular hyperparameter or a particular combination of hyperparameters, the deep learning model may be retrained according to one or more hyperparameters that are changed from the particular hyperparameter(s).

Further with respect to the option to reconfigure the deep learning model, the deep learning model may be updated with a layer substitution. For example, where the deep learning model included multiple particular layers, the deep learning model may be updated to include, replace, etc. one or more layers that are different from the particular layers. Similarly, the deep learning model may be updated with layer fusing (e.g. combining two or more of the particular layers).

Also with respect to the option to reconfigure the deep learning model, the deep learning model may be updated to use input stacking. For example, particular inputs last used by the deep learning model may be changed, such as by stacking inputs. The stacked inputs may be used to artificially increase the feature counts of tensors in the deep learning model. In other embodiments with respect to the option to reconfigure the deep learning model, the deep learning model may be updated to include changed code, such as high-level code (at a software level), or low level code (e.g. at a GPU level with GPU assembler code, or even machine code).

As noted above, any aspect(s) of the deep learning model, such as any combination of the embodiments mentioned above, may be updated to create the improved version of the deep learning model. As an option, the aspect(s) that are changed for updating the deep learning model may be selected automatically. For example, the aspect(s) may be iteratively changed until the improved deep learning model is generated.

With respect to the present description, the deep learning model may be considered to be improved from the last (or any prior) version of the deep learning model when any aspect, or any preselected aspect(s), of the deep learning model has improved, such as accuracy (e.g. ability to provide more accurate inferences which may improve an end-user experience), quality (e.g. quality of inferences), performance (e.g. improved speed, reduced resource consumption, etc.), etc. A version of the deep learning model resulting from any iteration of retraining may be considered “improved” when any improvement benchmark, or any preselected improvement benchmark(s), are met. The improvement benchmarks may be predefined (e.g. manually), for example as thresholds for each category of improvement (i.e. accuracy, quality, and/or performance) or even sub-category of improvement (e.g. improved speed, reduced resource consumption). As an option, improvement metrics may be measured when the updated deep learning model is executed by different CPUs or GPUs, in which case the updated deep learning model may be considered “improved” for only those CPUs and/or GPUs that enabled the updated deep learning model to meet the improvement benchmark(s).

In operation 203, a client with a previous version of the deep learning model is determined. The previous version of the deep learning model may refer to any version of the deep learning model generated prior to the updated version of the deep learning model generated in operation 202.

Once the deep learning model is updated to create the improved version of the deep learning model and the client with the previous version of the deep learning model is determined, the updated version of the deep learning model is automatically distributed to the client when the updated version of the deep learning model meets or exceeds one or more improvement benchmarks, as shown in operation 204). The client may be client 103 of FIG. 1, for example. In the embodiment described above where the updated deep learning model is considered “improved” for only certain CPUs and/or GPUs (i.e. that enabled the updated deep learning model to meet the improvement benchmark(s)), the improved version of the deep learning model may only be distributed to the client when the client includes one or more of those certain CPUs and/or GPUs. This may help ensure that the client is configured to be able to realize the improvements when executing the improved version of the deep learning model.

In one embodiment, the improved version of the deep learning model may be distributed to the client by communicating a copy of the improved version of the deep learning model to the client. To this end, the client may locally store, and thus locally execute, the copy of the improved version of the deep learning model. It should be noted that while the present method 200 references distributing the improved version of the deep learning model to a particular client, the method 200 may be implemented in other embodiments to distribute the improved version of the deep learning model to multiple different clients (e.g. that each have a previous version of the deep learning model).

It should be further noted that the improved version of the deep learning model may be distributed to the client responsive to a particular trigger. In one embodiment, the trigger may be the creation of the improved version of the deep learning model. In another embodiment, the trigger may be a scheduled distribution. In yet another embodiment, the trigger may be a request received by the client for an improved version of the deep learning model (e.g. as described in more detail below). When the server determines, responsive to the request, that it has a version of the deep learning model that has been updated from a version currently stored on the client, the server may distribute the updated version of the deep learning model to the client.

To this end, the method 200 may be implemented for the deep learning model for creating an improved version of the deep learning model that can be used by the client to perform inferencing operations. Whereas current optimizations of deep learning models typically involve software engineers or data scientists conducting experiments to find better solutions, the present method 200 may allow the server to attempt huge numbers of different possible combinations of changes to find improvements. This method 200 may be repeated over and over to provide ongoing and continuous deep learning model improvements that are then downloaded to the client to improve operations involving the deep learning model. Similarly, the method 200 may be implemented for other deep learning models to create improved versions of those deep learning models that can be used by any number of different clients to provide other types of inferenced data.

FIG. 3 illustrates a flowchart of a client method 300 for implementing an improved deep learning model that provides inferenced data to a local software application, in accordance with an embodiment. In one embodiment, the method 300 may be performed by the client 103 of FIG. 1.

In operation 301, a deep learning model is stored. In the context of the present method 300, the deep learning model is usable for providing inferenced data to a software application (e.g. such as the deep learning model 102 used by the software application 104 of FIG. 1). The deep learning model may be stored locally (e.g. by the client 103). In one embodiment, the deep learning model may be stored in a local repository with other deep learning models usable for providing other types of inferenced data to the software application or other software applications.

In operation 302, the deep learning model is executed to perform inferencing operations and to provide inferenced data to a software application. The deep learning model and the software application may both execute locally. In particular, the software application provides input data to the deep learning model which processes the input data to generate one or more inferences (i.e. inferenced data) for the input data. The inferenced data is output by the deep learning model to the software application for use by functions, tasks, etc. of the software application.

It should be noted that the software application may use the deep learning model as often as required while the deep learning model is stored and is thus accessible to the software application. For example, various functions within the software application, or multiple executions of the same function, may cause input data to be provided to the deep learning model for the purpose of obtaining the inferenced data.

In operation 303, an updated version of the deep learning model is received. Thus, after some period in which the deep learning model is executed to provide inferenced data to a software application, the improved version of the deep learning model may be received. In one embodiment, the improved version of the deep learning model may be received by a server (e.g. server 101 of FIG. 1).

As an option, the improved version of the deep learning model may be received responsive to a trigger. In one embodiment, the trigger may occur on the server side, and thus the improved version of the deep learning model may be provided to the client proactively. For example, the trigger may be the creation of the improved version of the deep learning model at the server. As another example, the trigger may be a scheduled distribution at the server.

In another embodiment, the trigger may occur on the client side. The trigger may be scheduled, may be the initiated execution of the software application that uses the deep learning model, or may be a call to a feature API that causes execution of the deep learning model. Responsive to the client-side trigger, the client may request from the server an improved version of the deep learning model. When the server determines, responsive to the request, that it has a version of the deep learning model that is updated from a version currently stored on the client, the server may distribute the updated version of the deep learning model to the client.

Further, in operation 304, the updated version of the deep learning model is executed to provide additional inferenced data to the software application. In one embodiment, the updated version of the deep learning model may replace the last version of the deep learning model used by the software application (i.e. in operation 302). To this end, the software application may use the updated version of the deep learning model once received by the client.

FIG. 4A illustrates a block diagram of a system 400 for updating a deep learning model that performs inferencing operations and provides inferenced data to a software application, in accordance with an embodiment. It should be noted that the definitions and/or descriptions provided with respect to the embodiments above may equally apply to the present description.

As shown, a client 401 has installed thereon a software application 402 that uses one or more deep learning models stored in a local deep learning model store 403. Each of the deep learning models may perform a different type of inferences and provide a different type of inferenced data, and thus may be usable (e.g. by the software application 402 and/or other software applications installed on the client) to obtain any needed inferenced data.

Additionally, a server 409 operates to update a deep learning model to create an updated version of the deep learning model 410. As shown, the server receives research data 404 which includes a new training dataset 407 and/or a new deep learning model design 408 (reconfiguration). The research data 404 may be generated from a newly generated public dataset 405 and/or from offline information 406 received in association with the software application.

The server 409 may update the deep learning model using manual training and tuning of the deep learning model by one or more users, and/or using automatic training and optimizing of the deep learning model by a neural network optimizer (not shown). The updated version of the deep learning model 410 is then distributed to the client 401 via a deep learning model update server 412 of a cloud service 411. Optionally, the client 401 may subscribe to the cloud service 411 to be provided access to deep learning models.

Each time the server 409 starts a new deep learning model training session, the metadata that describes the deep learning model, including all training hyperparameters, inferencing parameters, and the dataset, is stored either in a file or a database. This allows the deep learning model to be fully recreated at any time in the future. The server 409 can also use that metadata to conduct future experiments and to derive new deep learning models. At any point, the server 409 will likely have multiple deep learning models being trained and evaluated against improvement benchmarks.

FIG. 4B illustrates a flowchart of the method of the client 401 of FIG. 4, in accordance with an embodiment. As a first sub-process of the method of the client 401, during runtime of the software application 402 (operation 450), a feature API is invoked (operation 451). The feature API may provide an interface to the deep learning model to allow the software application 402 to interface with the deep learning model.

Responsive to the invocation of the feature API, the client 401 determines whether the deep learning model has been updated since a last call made to the deep learning model by the software application 402 (decision 452). The client 401 may accomplish this by querying the local deep learning model store 403 for a latest stored version of the deep learning model.

Responsive to determining that the deep learning model has not been updated, the client 401 runs the deep learning model (operation 454) with input data provided by the software application 402, and returns inferenced data output by the deep learning model back to the software application 402 (operation 455).

Responsive to determining that the deep learning model has been updated, the client 401 loads the updated (improved) deep learning model from the local deep learning model store 403 (operation 453). This may be performed as a hot-swap (in real-time) during execution of the software application. The client 401 then runs the updated deep learning model (operation 454) with input data provided by the software application 402, and returns inferenced data output by the deep learning model back to the software application 402 (operation 455).

As a second sub-process of the method of the client 401, the client 401 is triggered (operation 456) to check for a deep learning model update (operation 457). The trigger may be caused by a schedule, or by the feature API invocation in operation 451. The client 401 sends a request to the cloud service 411 for any updated version of the deep learning model. When the cloud service 411 has access to an updated version of the deep learning model, the client 401 downloads the updated deep learning model (operation 458) and stores the new model in the local deep learning model store 403 (operation 459).

Machine Learning

Deep neural networks (DNNs), including deep learning models, developed on processors have been used for diverse use cases, from self-driving cars to faster drug development, from automatic image captioning in online image databases to smart real-time language translation in video chat applications. Deep learning is a technique that models the neural learning process of the human brain, continually learning, continually getting smarter, and delivering more accurate results more quickly over time. A child is initially taught by an adult to correctly identify and classify various shapes, eventually being able to identify shapes without any coaching. Similarly, a deep learning or neural learning system needs to be trained in object recognition and classification for it get smarter and more efficient at identifying basic objects, occluded objects, etc., while also assigning context to objects.

At the simplest level, neurons in the human brain look at various inputs that are received, importance levels are assigned to each of these inputs, and output is passed on to other neurons to act upon. An artificial neuron or perceptron is the most basic model of a neural network. In one example, a perceptron may receive one or more inputs that represent various features of an object that the perceptron is being trained to recognize and classify, and each of these features is assigned a certain weight based on the importance of that feature in defining the shape of an object.

A deep neural network (DNN) model includes multiple layers of many connected nodes (e.g., perceptrons, Boltzmann machines, radial basis functions, convolutional layers, etc.) that can be trained with enormous amounts of input data to quickly solve complex problems with high accuracy. In one example, a first layer of the DNN model breaks down an input image of an automobile into various sections and looks for basic patterns such as lines and angles. The second layer assembles the lines to look for higher level patterns such as wheels, windshields, and mirrors. The next layer identifies the type of vehicle, and the final few layers generate a label for the input image, identifying the model of a specific automobile brand.

Once the DNN is trained, the DNN can be deployed and used to identify and classify objects or patterns in a process known as inference. Examples of inference (the process through which a DNN extracts useful information from a given input) include identifying handwritten numbers on checks deposited into ATM machines, identifying images of friends in photos, delivering movie recommendations to over fifty million users, identifying and classifying different types of automobiles, pedestrians, and road hazards in driverless cars, or translating human speech in real-time.

During training, data flows through the DNN in a forward propagation phase until a prediction is produced that indicates a label corresponding to the input. If the neural network does not correctly label the input, then errors between the correct label and the predicted label are analyzed, and the weights are adjusted for each feature during a backward propagation phase until the DNN correctly labels the input and other inputs in a training dataset. Training complex neural networks requires massive amounts of parallel computing performance, including floating-point multiplications and additions. Inferencing is less compute-intensive than training, being a latency-sensitive process where a trained neural network is applied to new inputs it has not seen before to classify images, translate speech, and generally infer new information.

Inference and Training Logic

As noted above, a deep learning or neural learning system needs to be trained to generate inferences from input data. Details regarding inference and/or training logic 515 for a deep learning or neural learning system are provided below in conjunction with FIGS. 5A and/or 5B.

In at least one embodiment, inference and/or training logic 515 may include, without limitation, a data storage 501 to store forward and/or output weight and/or input/output data corresponding to neurons or layers of a neural network trained and/or used for inferencing in aspects of one or more embodiments. In at least one embodiment data storage 501 stores weight parameters and/or input/output data of each layer of a neural network trained or used in conjunction with one or more embodiments during forward propagation of input/output data and/or weight parameters during training and/or inferencing using aspects of one or more embodiments. In at least one embodiment, any portion of data storage 501 may be included with other on-chip or off-chip data storage, including a processor's L1, L2, or L3 cache or system memory.

In at least one embodiment, any portion of data storage 501 may be internal or external to one or more processors or other hardware logic devices or circuits. In at least one embodiment, data storage 501 may be cache memory, dynamic randomly addressable memory (“DRAM”), static randomly addressable memory (“SRAM”), non-volatile memory (e.g., Flash memory), or other storage. In at least one embodiment, choice of whether data storage 501 is internal or external to a processor, for example, or comprised of DRAM, SRAM, Flash or some other storage type may depend on available storage on-chip versus off-chip, latency requirements of training and/or inferencing functions being performed, batch size of data used in inferencing and/or training of a neural network, or some combination of these factors.

In at least one embodiment, inference and/or training logic 515 may include, without limitation, a data storage 505 to store backward and/or output weight and/or input/output data corresponding to neurons or layers of a neural network trained and/or used for inferencing in aspects of one or more embodiments. In at least one embodiment, data storage 505 stores weight parameters and/or input/output data of each layer of a neural network trained or used in conjunction with one or more embodiments during backward propagation of input/output data and/or weight parameters during training and/or inferencing using aspects of one or more embodiments. In at least one embodiment, any portion of data storage 505 may be included with other on-chip or off-chip data storage, including a processor's L1, L2, or L3 cache or system memory. In at least one embodiment, any portion of data storage 505 may be internal or external to on one or more processors or other hardware logic devices or circuits. In at least one embodiment, data storage 505 may be cache memory, DRAM, SRAM, non-volatile memory (e.g., Flash memory), or other storage. In at least one embodiment, choice of whether data storage 505 is internal or external to a processor, for example, or comprised of DRAM, SRAM, Flash or some other storage type may depend on available storage on-chip versus off-chip, latency requirements of training and/or inferencing functions being performed, batch size of data used in inferencing and/or training of a neural network, or some combination of these factors.

In at least one embodiment, data storage 501 and data storage 505 may be separate storage structures. In at least one embodiment, data storage 501 and data storage 505 may be same storage structure. In at least one embodiment, data storage 501 and data storage 505 may be partially same storage structure and partially separate storage structures. In at least one embodiment, any portion of data storage 501 and data storage 505 may be included with other on-chip or off-chip data storage, including a processor's L1, L2, or L3 cache or system memory.

In at least one embodiment, inference and/or training logic 515 may include, without limitation, one or more arithmetic logic unit(s) (“ALU(s)”) 510 to perform logical and/or mathematical operations based, at least in part on, or indicated by, training and/or inference code, result of which may result in activations (e.g., output values from layers or neurons within a neural network) stored in an activation storage 520 that are functions of input/output and/or weight parameter data stored in data storage 501 and/or data storage 505. In at least one embodiment, activations stored in activation storage 520 are generated according to linear algebraic and or matrix-based mathematics performed by ALU(s) 510 in response to performing instructions or other code, wherein weight values stored in data storage 505 and/or data 501 are used as operands along with other values, such as bias values, gradient information, momentum values, or other parameters or hyperparameters, any or all of which may be stored in data storage 505 or data storage 501 or another storage on or off-chip. In at least one embodiment, ALU(s) 510 are included within one or more processors or other hardware logic devices or circuits, whereas in another embodiment, ALU(s) 510 may be external to a processor or other hardware logic device or circuit that uses them (e.g., a co-processor). In at least one embodiment, ALUs 510 may be included within a processor's execution units or otherwise within a bank of ALUs accessible by a processor's execution units either within same processor or distributed between different processors of different types (e.g., central processing units, graphics processing units, fixed function units, etc.). In at least one embodiment, data storage 501, data storage 505, and activation storage 520 may be on same processor or other hardware logic device or circuit, whereas in another embodiment, they may be in different processors or other hardware logic devices or circuits, or some combination of same and different processors or other hardware logic devices or circuits. In at least one embodiment, any portion of activation storage 520 may be included with other on-chip or off-chip data storage, including a processor's L1, L2, or L3 cache or system memory. Furthermore, inferencing and/or training code may be stored with other code accessible to a processor or other hardware logic or circuit and fetched and/or processed using a processor's fetch, decode, scheduling, execution, retirement and/or other logical circuits.

In at least one embodiment, activation storage 520 may be cache memory, DRAM, SRAM, non-volatile memory (e.g., Flash memory), or other storage. In at least one embodiment, activation storage 520 may be completely or partially within or external to one or more processors or other logical circuits. In at least one embodiment, choice of whether activation storage 520 is internal or external to a processor, for example, or comprised of DRAM, SRAM, Flash or some other storage type may depend on available storage on-chip versus off-chip, latency requirements of training and/or inferencing functions being performed, batch size of data used in inferencing and/or training of a neural network, or some combination of these factors. In at least one embodiment, inference and/or training logic 515 illustrated in FIG. 5A may be used in conjunction with an application-specific integrated circuit (“ASIC”), such as Tensorflow® Processing Unit from Google, an inference processing unit (IPU) from Graphcore™, or a Nervana® (e.g., “Lake Crest”) processor from Intel Corp. In at least one embodiment, inference and/or training logic 515 illustrated in FIG. 5.A may be used in conjunction with central processing unit (“CPU”) hardware, graphics processing unit (“GPU”) hardware or other hardware, such as field programmable gate arrays (“FPGAs”).

FIG. 5B illustrates inference and/or training logic 515, according to at least one embodiment. In at least one embodiment, inference and/or training logic 515 may include, without limitation, hardware logic in which computational resources are dedicated or otherwise exclusively used in conjunction with weight values or other information corresponding to one or more layers of neurons within a neural network. In at least one embodiment, inference and/or training logic 515 illustrated in FIG. 5.B may be used in conjunction with an application-specific integrated circuit (ASIC), such as Tensorflow® Processing Unit from Google, an inference processing unit (IPU) from Graphcore™, or a Nervana® (e.g., “Lake Crest”) processor from Intel Corp. In at least one embodiment, inference and/or training logic 515 illustrated in FIG. 5.B may be used in conjunction with central processing unit (CPU) hardware, graphics processing unit (GPU) hardware or other hardware, such as field programmable gate arrays (FPGAs). In at least one embodiment, inference and/or training logic 515 includes, without limitation, data storage 501 and data storage 505, which may be used to store weight values and/or other information, including bias values, gradient information, momentum values, and/or other parameter or hyperparameter information. In at least one embodiment illustrated in FIG. 5B, each of data storage 501 and data storage 505 is associated with a dedicated computational resource, such as computational hardware 502 and computational hardware 506, respectively. In at least one embodiment, each of computational hardware 506 comprises one or more ALUs that perform mathematical functions, such as linear algebraic functions, only on information stored in data storage 501 and data storage 505, respectively, result of which is stored in activation storage 520.

In at least one embodiment, each of data storage 501 and 505 and corresponding computational hardware 502 and 506, respectively, correspond to different layers of a neural network, such that resulting activation from one “storage/computational pair 501/502” of data storage 501 and computational hardware 502 is provided as an input to next “storage/computational pair 505/506” of data storage 505 and computational hardware 506, in order to mirror conceptual organization of a neural network. In at least one embodiment, each of storage/computational pairs 501/502 and 505/506 may correspond to more than one neural network layer. In at least one embodiment, additional storage/computation pairs (not shown) subsequent to or in parallel with storage computation pairs 501/502 and 505/506 may be included in inference and/or training logic 515.

Neural Network Training and Deployment

FIG. 6 illustrates another embodiment for training and deployment of a deep neural network. In at least one embodiment, untrained neural network 606 is trained using a training dataset 602. In at least one embodiment, training framework 604 is a PyTorch framework, whereas in other embodiments, training framework 604 is a Tensorflow, Boost, Caffe, Microsoft Cognitive Toolkit/CNTK, MXNet, Chainer, Keras, Deeplearning4j, or other training framework. In at least one embodiment training framework 604 trains an untrained neural network 606 and enables it to be trained using processing resources described herein to generate a trained neural network 608. In at least one embodiment, weights may be chosen randomly or by pre-training using a deep belief network. In at least one embodiment, training may be performed in either a supervised, partially supervised, or unsupervised manner.

In at least one embodiment, untrained neural network 606 is trained using supervised learning, wherein training dataset 602 includes an input paired with a desired output for an input, or where training dataset 602 includes input having known output and the output of the neural network is manually graded. In at least one embodiment, untrained neural network 606 is trained in a supervised manner processes inputs from training dataset 602 and compares resulting outputs against a set of expected or desired outputs. In at least one embodiment, errors are then propagated back through untrained neural network 606. In at least one embodiment, training framework 604 adjusts weights that control untrained neural network 606. In at least one embodiment, training framework 604 includes tools to monitor how well untrained neural network 606 is converging towards a model, such as trained neural network 608, suitable to generating correct answers, such as in result 614, based on known input data, such as new data 612. In at least one embodiment, training framework 604 trains untrained neural network 606 repeatedly while adjust weights to refine an output of untrained neural network 606 using a loss function and adjustment algorithm, such as stochastic gradient descent. In at least one embodiment, training framework 604 trains untrained neural network 606 until untrained neural network 606 achieves a desired accuracy. In at least one embodiment, trained neural network 608 can then be deployed to implement any number of machine learning operations.

In at least one embodiment, untrained neural network 606 is trained using unsupervised learning, wherein untrained neural network 606 attempts to train itself using unlabeled data. In at least one embodiment, unsupervised learning training dataset 602 will include input data without any associated output data or “ground truth” data. In at least one embodiment, untrained neural network 606 can learn groupings within training dataset 602 and can determine how individual inputs are related to untrained dataset 602. In at least one embodiment, unsupervised training can be used to generate a self-organizing map, which is a type of trained neural network 608 capable of performing operations useful in reducing dimensionality of new data 612. In at least one embodiment, unsupervised training can also be used to perform anomaly detection, which allows identification of data points in a new dataset 612 that deviate from normal patterns of new dataset 612.

In at least one embodiment, semi-supervised learning may be used, which is a technique in which in training dataset 602 includes a mix of labeled and unlabeled data. In at least one embodiment, training framework 604 may be used to perform incremental learning, such as through transferred learning techniques. In at least one embodiment, incremental learning enables trained neural network 608 to adapt to new data 612 without forgetting knowledge instilled within network during initial training.

Data Center

FIG. 7 illustrates an example data center 700, in which at least one embodiment may be used. In at least one embodiment, data center 700 includes a data center infrastructure layer 710, a framework layer 720, a software layer 730 and an application layer 740.

In at least one embodiment, as shown in FIG. 7, data center infrastructure layer 710 may include a resource orchestrator 712, grouped computing resources 714, and node computing resources (“node C.R.s”) 716(1)-716(N), where “N” represents any whole, positive integer. In at least one embodiment, node C.R.s 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, 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 cooling modules, etc. In at least one embodiment, one or more node C.R.s from among node C.R.s 716(1)-716(N) may be a server having one or more of above-mentioned computing resources.

In at least one embodiment, grouped computing resources 714 may include separate groupings of node C.R.s housed within one or more racks (not shown), or many racks housed in data centers at various geographical locations (also not shown). separate groupings of node C.R.s within grouped computing resources 714 may include grouped compute, network, memory or storage resources that may be configured or allocated to support one or more workloads. In at least one embodiment, several node C.R.s including CPUs or processors may grouped within one or more racks to provide compute resources to support one or more workloads. In at least one embodiment, one or more racks may also include any number of power modules, cooling modules, and network switches, in any combination.

In at least one embodiment, resource orchestrator 722 may configure or otherwise control one or more node C.R.s 716(1)-716(N) and/or grouped computing resources 714. In at least one embodiment, resource orchestrator 722 may include a software design infrastructure (“SDI”) management entity for data center 700. In at least one embodiment, resource orchestrator may include hardware, software or some combination thereof.

In at least one embodiment, as shown in FIG. 7, framework layer 720 includes a job scheduler 732, a configuration manager 734, a resource manager 736 and a distributed file system 738. In at least one embodiment, framework layer 720 may include a framework to support software 732 of software layer 730 and/or one or more application(s) 742 of application layer 740. In at least one embodiment, software 732 or application(s) 742 may respectively include web-based service software or applications, such as those provided by Amazon Web Services, Google Cloud and Microsoft Azure. In at least one embodiment, framework layer 720 may be, but is not limited to, a type of free and open-source software web application framework such as Apache Spark™ (hereinafter “Spark”) that may utilize 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 layers of data center 700. In at least one embodiment, configuration manager 734 may be capable of configuring different layers such as software layer 730 and framework layer 720 including Spark and distributed file system 738 for supporting large-scale data processing. In at least one embodiment, resource manager 736 may be capable of managing clustered or grouped computing resources mapped to or allocated for support of distributed file system 738 and job scheduler 732. In at least one embodiment, clustered or grouped computing resources may include grouped computing resource 714 at data center infrastructure layer 710. In at least one embodiment, resource manager 736 may coordinate with resource orchestrator 712 to manage these mapped or allocated computing resources.

In at least one embodiment, software 732 included in software layer 730 may include software used by at least portions of node C.R.s 716(1)-716(N), grouped computing resources 714, and/or distributed file system 738 of framework layer 720. one or more types of software may include, but are not limited to, Internet web page search software, e-mail virus scan software, database software, and streaming video content software.

In at least one embodiment, application(s) 742 included in application layer 740 may include one or more types of applications used by at least portions of node C.R.s 716(1)-716(N), grouped computing resources 714, and/or distributed file system 738 of framework layer 720. one or more types of applications may include, but are not limited to, any number of a genomics application, a cognitive compute, and a machine learning application, including training or inferencing software, machine learning framework software (e.g., PyTorch, TensorFlow, Caffe, etc.) or other machine learning applications used in conjunction with one or more embodiments.

In at least one embodiment, any of configuration manager 734, resource manager 736, and resource orchestrator 712 may implement any number and type of self-modifying actions based on any amount and type of data acquired in any technically feasible fashion. In at least one embodiment, self-modifying actions may relieve a data center operator of data center 700 from making possibly bad configuration decisions and possibly avoiding underutilized and/or poor performing portions of a data center.

In at least one embodiment, data center 700 may include tools, services, software or other resources to train one or more machine learning models or predict or infer information using one or more machine learning models according to one or more embodiments described herein. For example, in at least one embodiment, a machine learning model may be trained by calculating weight parameters according to a neural network architecture using software and computing resources described above with respect to data center 700. In at least one embodiment, trained machine learning models corresponding to one or more neural networks may be used to infer or predict information using resources described above with respect to data center 700 by using weight parameters calculated through one or more training techniques described herein.

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

Inference and/or training logic 515 are used to perform inferencing and/or training operations associated with one or more embodiments. In at least one embodiment, inference and/or training logic 515 may be used in system FIG. 7 for inferencing or predicting operations based, at least in part, on weight parameters calculated using neural network training operations, neural network functions and/or architectures, or neural network use cases described herein.

As described herein, a method, computer readable medium, and system are disclosed for improving deep learning models that perform inferencing operations to provide inferenced data to software applications. In accordance with FIGS. 1-4B, an embodiment may provide a deep learning model usable for performing inferencing operations and for providing inferenced data, where the deep learning model is stored (partially or wholly) in one or both of data storage 501 and 505 in inference and/or training logic 515 as depicted in FIGS. 5A and 5B. Training and deployment of the deep learning model may be performed as depicted in FIG. 6 and described herein. For example, the deep learning model, when untrained, may subsequently be trained using training framework 604. Additionally, the deep learning model, when previously trained, may be updated to create an updated version of the deep learning model also using framework 604. Further, the updated version of a deep learning model may be distributed to a client for use in providing the inferenced data. Distribution of the trained or re-trained deep learning model may be performed using one or more servers in a data center 700 as depicted in FIG. 7 and described herein.

Claims

1. A method, comprising:

storing a deep learning model usable to perform inferencing operations and for providing inferenced data;

receiving one or more updates to the deep learning model;

updating the deep learning model to create an updated version of the deep learning model;

determining a client computing device with a previous version of the deep learning model; and

automatically distributing one or more updates to the deep learning model to the client computing device when the updated version of the deep learning model meets or exceeds one or more improvement benchmarks.

2. The method of claim 1, wherein the deep learning model is usable to perform the inferencing operations and to output the inferenced data to a software application installed on the client computing device.

3. The method of claim 2, wherein the software application is a video game.

4. The method of claim 1, wherein the updating the deep learning model comprises retraining the deep learning model.

5. The method of claim 5, wherein the retraining the deep learning model comprises retraining the deep learning model using a changed dataset.

6. The method of claim 1, wherein the automatically distributing the one or more updates to the deep learning model comprises distributing one or more updated parameters for the previous version of the deep learning model to the client computing device.

7. The method of claim 6, wherein the automatically distributing the one or more updates to the deep learning model comprises distributing one or more updated hyperparameter adjustments for the previous version of the deep learning model to the client computing device.

8. The method of claim 6, wherein the automatically distributing the one or more updates to the deep learning model comprises updating the previous version of the deep learning model of the client computing device with at least one of a layer substitution, a layer fusing, or input stacking.

9. The method of claim 1, wherein the updating is performed automatically by software or other neural networks.

10. The method of claim 1, wherein the one or more improvement benchmarks are one or more thresholds for improvement relating to accuracy, quality, or performance.

11. A non-transitory computer-readable medium storing computer instructions that, when executed by one or more processors, cause the one or more processors to perform a method comprising:

storing a deep learning model;

executing the deep learning model to perform inferencing operations and to provide inferenced data to a software application;

receiving an updated version of the deep learning model; and

executing the updated version of the deep learning model to provide additional inferenced data to the software application.

12. The non-transitory computer-readable medium of claim 11, wherein the deep learning model and the software application are installed on a client computing device.

13. The non-transitory computer-readable medium of claim 11, wherein the software application is a video game.

14. The non-transitory computer-readable medium of claim 13, wherein the inferenced data includes one or more image-related inferences, the one or more image-related inferences being at least one of an anti-aliased image, an image with upscaled resolution, or a denoised image.

15. The non-transitory computer-readable medium of claim 11, wherein the software application is a voice recognition application.

16. The non-transitory computer-readable medium of claim 15, wherein the inferenced data includes one or more audio-related inferences, the one or more audio-related inferences being at least one of a language translation or a voice recognized command.

17. The non-transitory computer-readable medium of claim 11, wherein executing the deep learning model to perform inferencing operations and to provide inferenced data to a software application includes:

providing input data by the software application to the deep learning model,

processing the input data by the deep learning model to generate the inferenced data, and

outputting the inferenced data to the software application.

18. The non-transitory computer-readable medium of claim 11, wherein the updated version of the deep learning model is received responsive to a trigger, the trigger being a request for the updated version of the deep learning model being sent to a server.

19. The non-transitory computer-readable medium of claim 18, wherein the request is triggered as a result of a call to a feature application programming interface (API) that causes execution of the deep learning model.

20. A system, comprising:

a memory storing instructions; and

one or more processors that execute the instructions to perform a method comprising: receiving a deep learning model from a server; storing the deep learning model; executing the deep learning model to perform inferencing operations and to provide inferenced data to a software application; sending a request to the server to determine whether an updated version of the deep learning model is available; receiving the updated version of the deep learning model when an updated version of the deep learning model is available that meets or exceeds one or more improvement thresholds; and executing the updated version of the deep learning model to provide additional inferenced data to the software application.