Private training with server-side forward pass

Abstract

Systems and methods are provided. A method can include providing, by a first computing device to a computing system comprising one or more second computing devices, data indicative of one or more input values for one or more layers of the machine-learned model. The method can include receiving, by the first computing device from the computing system, data indicative of a forward pass of the machine-learned model based on the one or more input values. In the method, the data indicative of the forward pass can include a reduced-parameter-count model description associated with the machine-learned model. The method can include determining, by the first computing device based on the data indicative of the forward pass, one or more update values for the machine-learned model.

Description

PRIVATE TRAINING WITH SERVER-SIDE FORWARD PASSFIELD

[0001] The present disclosure relates generally to machine learning processes and machine-learned devices and systems. More particularly, the present disclosure relates to data-security-preserving systems and methods for training a machine-learned model using secure, confidential, or otherwise access-restricted training data.BACKGROUND

[0002] A computer can receive input(s). The computer can execute instructions to process the input(s) to generate output(s) using a parameterized model. The computer can obtain feedback on its performance in generating the outputs with the model. The computer can generate feedback by evaluating its performance. The computer can receive feedback from an external source. The computer can update parameters of the model based on the feedback to improve its performance. In this manner, the computer can iteratively “learn” to generate the desired outputs. The resulting model is often referred to as a machine-learned model.SUMMARY

[0003] Aspects and advantages of embodiments of the present disclosure will be set forth in part in the following description, or can be learned from the description, or can be learned through practice of the embodiments.

[0004] Example aspects of the present disclosure provide an example method. In some implementations, the example method can include providing, by a first computing device to a computing system comprising one or more second computing devices, data indicative of one or more input values for one or more layers of a machine-learned model. The example method can include receiving, by the first computing device from the computing system, data indicative of a forward pass of the machine-learned model based on the one or more input values. In the example method, the data indicative of the forward pass can include a reduced-parameter-count model description associated with the machine- learned model. The example method can include determining, by the first computing device based on the data indicative of the forward pass, one or more update values for the machine-learned model. In the example method, the one or more update values can include at least one update value for at least one parameter of the reduced-parameter-count model description.

[0005] The example method can include providing, by the first computing device, the one or more update values to the computing system.

[0006] The example method can include updating, by the first computing device based on the one or more update values, one or more local parameters stored locally on the first computing device, wherein the local parameters are associated with the machine-learned model.

[0007] In the example method, the first computing device can be a first client device. In some implementations, the computing system can include one or more server devices. In some implementations, the forward pass can have been performed by the one or more server devices.

[0008] In the example method, the computing system can further include one or more second client devices. In some implementations, the example method can further include performing, by the first computing device, at least one of: masking the one or more update values and subsequently providing a result of the masking to the one or more second client devices; and receiving at least one masked update value from the one or more second client devices and providing one or more aggregated update values to the computing system based on the one or more update values and the at least one masked update value.

[0009] In the example method, the data indicative of the forward pass can include data indicative of a strict subset of a plurality of parameters of a layer of the one or more layers.

[0010] In the example method, the strict subset can consist of less than 50 percent of the plurality of parameters of the layer of the one or more layers.

[0011] In the example method, the strict subset can consist of less than 10 percent of the plurality of parameters of the layer of the one or more layers.

[0012] In the example method, the reduced-parameter-count model description can include one or more first parameters of the one or more layers, the first parameters having a non-zero-valued contribution to one or more output values of a final layer of the one or more layers during the forward pass. In the example method, the reduced-parameter-count model description can omit one or more second parameters of the one or more layers, the second parameters having a zero-valued contribution to the one or more output values during the forward pass.

[0013] In the example method, determining the one or more update values can include computing, based at least in part on the one or more first parameters, a gradient of an objective function associated with the machine-learned model. In some implementations of the example method, the objective function can be based at least in part on the one or more output values.

[0014] In the example method, the data indicative of the forward pass can include a plurality of output activation values of a plurality of nodes of the one or more layers.

[0015] In the example method, data indicative of the one or more input values can include data indicative of a user input received from a user of the first computing device, the user input comprising one or more of text-based input, audio input, image input, and video input.

[0016] In the example method, the one or more update values can be determined based on private data not available to the computing system.

[0017] In the example method, the machine-learned model can include a sequence processing model.

[0018] Example aspects of the present disclosure provide one or more example non- transitory computer-readable media storing instructions that are executable by one or more processors to cause a computing system to perform example operations. In some implementations, the example operations can include receiving, from a first computing device, data indicative of one or more input values for one or more layers of the machine- learned model. The example operations can include performing a forward pass of the machine-learned model based on the data indicative of the one or more input values. The example operations can include providing, to the first computing device, data indicative of the forward pass. In the example operations, the data indicative of the forward pass can include a reduced-parameter-count model description of the machine-learned model.

[0019] In some implementations, the example operations can include receiving, from the first computing device, data indicative of one or more update values for the machine- learned model. The example operations can include updating the machine-learned model based at least in part on the data indicative of the update values.

[0020] In the example operations, the data indicative of the forward pass can include data indicative of a strict subset of a plurality of parameters of a layer of the one or more layers.

[0021] In the example operations, the reduced-parameter-count model description can include one or more first parameters of the one or more layers, the first parameters having anon-zero-valued contribution to one or more output values of a final layer of the one or more layers during the forward pass. In the example operations, the reduced-parameter-count model description can omit one or more second parameters of the one or more layers, the second parameters having a zero-valued contribution to the one or more output values during the forward pass.

[0022] In the example operations, the data indicative of the forward pass can include a plurality of output activation values of a plurality of nodes of the one or more layers.

[0023] Example aspects of the present disclosure provide an example computing system that includes one or more processors and one or more example non-transitory computer-readable media storing instructions that are executable by one or more processors to cause a computing system to perform example operations. In some implementations, the example operations can include providing, to a second computing system comprising one or more computing devices, data indicative of one or more input values for one or more layers of a machine-learned model. The example operations can include receiving, from the second computing system, data indicative of a forward pass of the machine-learned model based on the one or more input values, wherein the data indicative of the forward pass comprises a reduced-parameter-count model description of the machine-learned model. The example operations can include determining, based on the data indicative of the forward pass, one or more update values for the machine-learned model. In the example operations, the one or more update values can include at least one update value for at least one parameter of the reduced-parameter-count model description.

[0024] Other example aspects of the present disclosure are directed to other systems, methods, apparatuses, tangible non-transitory computer-readable media, and devices for performing functions described herein. These and other features, aspects, and advantages of various implementations will become better understood with reference to the following description and appended claims. The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate implementations of the present disclosure and, together with the description, help explain the related principles.BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure l is a block diagram of an example system for training a machine learned model according to example aspects of the present disclosure;

[0026] Figure 2 is a block diagram of an example system for training a machine learned model according to example aspects of the present disclosure;

[0027] Figure 3 is a block diagram of an example system for training a machine learned model according to example aspects of the present disclosure;

[0028] Figure 4 is a block diagram of an example system for training a machine learned model according to example aspects of the present disclosure;

[0029] Figure 5A is a first block diagram of an example system for communicating model update values according to example aspects of the present disclosure;

[0030] Figure 5B is a second block diagram of an example system for communicating model update values according to example aspects of the present disclosure;

[0031] Figure 6 is a flow chart diagram illustrating an example method for training a machine-learned model according to example aspects of the present disclosure;

[0032] Figure 7 is a flow chart diagram illustrating an example method for training a machine-learned model according to example aspects of the present disclosure;

[0033] Figure 8 is a flow chart diagram illustrating an example method for training a machine-learned model according to example implementations of aspects of the present disclosure;

[0034] Figure 9 is a block diagram of an example processing flow for using machine- learned model(s) to process input(s) to generate output(s) according to example implementations of aspects of the present disclosure;

[0035] Figure 10 is a block diagram of an example sequence processing model according to example implementations of aspects of the present disclosure;

[0036] Figure 11 is a block diagram of an example technique for populating an example input sequence for processing by a sequence processing model according to example implementations of aspects of the present disclosure;

[0037] Figure 12 is a block diagram of an example model development platform according to example implementations of aspects of the present disclosure;

[0038] Figure 13 is a block diagram of an example training workflow for training a machine-learned model according to example implementations of aspects of the present disclosure;

[0039] Figure 14 is a block diagram of an inference system for operating one or more machine-learned model(s) to perform inference according to example implementations of aspects of the present disclosure;

[0040] Figure 15 is a block diagram of an example networked computing system according to example implementations of aspects of the present disclosure;

[0041] Figure 16 is a block diagram of an example computing device according to example implementations of aspects of the present disclosure; and

[0042] Figure 17 is a block diagram of an example computing device according to example implementations of aspects of the present disclosure.DETAILED DESCRIPTION

[0043] Generally, the present disclosure is directed to systems and methods for data- security-preserving training of a machine-learned model at a reduced client-side computational cost compared to some alternative methods. More particularly, a client-side device can provide, to a server device, one or more training inputs (e.g., non-private input values) associated with a training example. The server device can provide the training inputs to a machine-learned model and perform a forward pass using the machine-learned model. Based on the forward pass, the server device can provide, to the client device, a reduced- parameter-count model description (e.g., sparse model graph, etc.) associated with the machine-learned model, which may be equivalent to the full-parameter-count machine- learned model for the purposes of the training example. Based on the reduced-parametercount model description, the client device can determine one or more model updates (e.g., using gradient descent methods) for the machine-learned model based on private or secure training labels associated with the training example, wherein the secure training labels are not shared with the server device. In this manner, for instance, a client-side computational cost of determining a model update can be reduced compared to some alternative methods for data- security-preserving training using private, secure, or otherwise access-restricted client-side data.

[0044] For example, in some instances, the machine-learned model can include a machine-learned model characterized by sparse activation outputs, wherein a significant percentage (e.g., greater than 50 percent; greater than 90 percent; greater than 99 percent; etc.) of activation outputs (i.e., outputs of one or more nodes of the machine-learned model) may be equal to zero during some forward passes. In some such instances, nodes or weights associated with such zero-valued activation outputs may be irrelevant or safe to ignore when determining a model update according to some training methods. For example, in some training methods based on gradient descent, input weights to some nodes associated with a zero-valued output activation may be associated with a zero-valued partial derivative for all possible values of the private or secure training labels. In such instances, a reduced-parametermodel description can include a description of the weights and nodes that are necessary to compute a model update, and can omit any weights and nodes that are unnecessary to compute a model update based on the training example.

[0045] Determining one or more model updates based on forward pass data received from a server device can include, for example, gradient-based methods such as gradient descent. For example, a server device can provide forward pass data comprising a plurality of weights associated with the forward pass (e.g., weights having a non-zero impact on an output of the forward pass, etc.); a plurality of intermediate values or output values computed by the server during the forward pass (e.g., output activations of neural network nodes; matrix multiplication outputs; etc.); and other relevant model description data, such as index values identifying locations (e.g., node locations, edge locations, etc.) within the fullparameter-count machine-learned model associated with each weight, intermediate value, or the like. Based on the received forward pass data and a secure training label associated with the training example, the client device can compute a gradient associated with an objective function (e.g., loss function, etc.) used for training the machine-learned model, and can determine one or more update values based on the gradient (e.g., by multiplying the gradient by a learning rate parameter, aggregating the gradient with other gradients of a mini-batch, etc.).

[0046] After computing one or more model update values based on a reduced- parameter-count-model description received from a server device, a client device can store the update values locally or share them with one or more other devices (e.g., for aggregation of a plurality of updates from a plurality of client devices, etc.). For example, in some instances, systems and methods according to some aspects of the present disclosure can include federated reconstruction or partially local federated learning, wherein one or more machine-learned models may be partitioned into a global component (e.g., global model; one or more global layers; etc.) that is shared between devices, and one or more respective local components (e.g., local model, layers, parameters, etc.) that are stored locally on each client device. In such instances, the client device can store or use the update values locally, such as by updating the respective local component associated with the client device.

[0047] As another example, in some instances, the client can share (e.g., directly or indirectly) one or more update values with one or more other computing devices, such as one or more other client devices; the server device that performed the forward pass; or another server device. In some instances, the sharing can be performed in a manner that preserves the data security of private training labels used to compute the model updates. For example, insome instances, systems and methods according to some aspects of the present disclosure can include various forms of secure aggregation, differentially private aggregation, or the like. As non-limiting illustrative examples, a client device can mask (e.g., encrypt, add random noise to, split into a plurality of component update values, etc.) the one or more update values before sharing with one or more other devices (e.g., plurality of client devices, shuffler server device separate from the forward pass server device, etc.); perform peer-to-peer sharing with a plurality of client devices; aggregate a plurality of masked update values before providing the aggregated masked values to a server device; and the like. In this manner, for instance, data security can be preserved while performing model training at reduced client-side computational cost compared to some alternative methods of data-security-preserving federated learning.

[0048] In some instances, systems and methods according to example aspects of the present disclosure can include or be combined with additional methods for reducing a clientside computational cost of federated learning. For example, in some instances, systems and methods according to example aspects of the present disclosure can include parameterefficient federated learning, such as federated learning using parameter-efficient adapter methods (e.g., adapter-based fine-tuning, low-rank adaptation, etc.). For example, a model comprising a trainable adapter, wherein the adapter has a reduced parameter count compared to a full machine-learned model, can be trained by providing non-private input values to a server device; performing a forward pass using the server device; providing a reduced- parameter-count model description (e.g., comprising a reduced-parameter-count adapter description) to the client device; and determining one or more update values using the client device (e.g., using secure training labels that are not shared with the server device).

[0049] Systems and methods according to some aspects of the present disclosure can include training various machine-learned model architectures, such as neural network architectures (e.g., recursive neural networks, convolutional neural networks, transformers, or other architectures comprising one or more layers of weights or activations); sequence processing model architectures (e.g., language models, audio processing models, video processing models, etc.); recommender models (e.g., model for suggesting a movie, song, website, etc.); retrieval models (e.g., model for retrieving a document, website, media item, or other item based on an input query, etc.); high-sparsity model architectures (e.g., sparse mixture-of-expert models, etc.); or other machine-learned models.

[0050] Example fields of application for systems and methods according to some aspects of the present disclosure can include various fields of application wherein clientdevices may share input data (e.g., non-private input data) with a server device comprising a machine-learned model. By way of non-limiting illustrative example, systems and methods according to aspects of the present disclosure could be applied to machine-learned search (e.g., web search, etc.), wherein a user may provide a search query input to one or more server devices; machine-learned interactive applications (e.g., chatbots, machine-learned agents or digital assistants, etc.), wherein a user may provide an interactive input to one or more server devices; generative artificial intelligence (e.g., image generation, video generation, language generation, etc.), wherein a user may provide an input context or prompt to a generative machine-learned model; or the like.

[0051] In some instances, an example field of application can include a field of application where relevant non-input data (e.g., reinforcement learning feedback data, etc.) is not shared with the server (e.g., for reasons of security, data privacy, or other data access restrictions). As a non-limiting illustrative example, relevant non-input data can include one or more secure or private signals of machine-learned output quality known to the client device, such as an action taken or not taken by a user (e.g., opening or viewing a retrieved or recommended item; undoing or cancelling an action performed by a machine-learned agent; saving a generated image; etc.); a secure or private metric known to the client device (e.g., dwelling time on a recommended, searched, or retrieved web site, movie, or other item); or any other data associated with an objective function (e.g., loss function) for training a machine-learned model.

[0052] Systems and methods according to example aspects of the present disclosure can provide a variety of technical effects and benefits, such as improvements to the training of machine-learned models. For example, systems and methods according to example aspects of the present disclosure can train a machine-learned model at reduced computational cost; train larger or more complex models at a similar (e.g., same) computational cost; provide increased technical performance for a similar (e.g., same) computational cost; or provide increased data security for a given computational cost compared to some alternative implementations.

[0053] For example, systems and methods according to example aspects of the present disclosure can reduce a client-side computational cost (e.g., electricity cost, memory usage, processor usage, etc.) of training a machine-learned model in multiple ways. First, performing a forward pass on a server device can reduce a client-side computational cost (e.g., compared to alternative fully-client-side training methods) by eliminating any need to perform the forward pass on the client device. Second, by determining a reduced-parameter-count model description to provide to the client device, the server device can reduce the number of parameters for which a client device must compute an update value, partial derivative, or the like. For example, because computing a partial derivative or other value costs computational resources (e.g., processor resources, memory resources, electricity, etc.), reducing a number of partial derivatives computed on the client side (e.g., according to gradient-based training methods) can reduce an overall client-side computational cost of training the machine-learned model.

[0054] Additionally, in some instances, reducing a client-side computational cost can reduce an overall computational cost of training a machine-learned model due to server-side efficiencies. For example, in some instances, a dedicated server device may comprise high- efficiency dedicated hardware (e.g., application-specific integrated circuits such as tensor processing units, graphics processing units, etc.) that can perform a forward pass or other computation at reduced cost compared to a client device. As a non-limiting illustrative example, a server-side tensor processing unit (TPU) or graphics processing unit (GPU) may have a lower cost per floating-point operation (e.g., more floating-point operations per watt of electricity consumed, etc.); larger amounts of on-chip random access memory (RAM), leading to reduced computational cost of training large models (e.g., due to reduced communication overhead, etc.); and other hardware-based efficiency advantages over a client-side device (e.g., smartphone, laptop, desktop, tablet, etc.).

[0055] Additionally, in some instances, performing a forward pass on a server device can enable a single server-side forward pass to be used for a plurality of training iterations associated with a plurality of client devices (e.g., instead of a plurality of client-side forward passes), thereby reducing an overall computational cost of performing a plurality of training iterations by reducing the number of times the forward pass is performed. As a non-limiting illustrative example, a server device may receive a plurality of similar (e.g., identical) input values (e.g., identical search input values such as “things to do in Cleveland”; identical chatbot input values such as “hi”; etc.) from a plurality of client devices. The plurality of client devices may store a plurality of training labels (e.g., private or secure training labels) associated with the similar (e.g., same) inputs, and the training labels may be very different from each other (e.g., different websites visited in response to similar search results, one user clicking “thumbs up” and another user clicking “thumbs down” in response to identical chatbot outputs, etc.). In some instances, the server device may store already-computed forward pass data and provide the same forward pass data to a plurality of different client devices associated with similar (e.g., same) inputs, thereby reducing a number of forwardpasses performed and reducing a computational cost of performing a plurality of training iterations based on a plurality of training examples with similar (e.g., identical) training inputs and different (e.g., user-dependent, etc.) private training labels.

[0056] Additionally, in some instances, reducing a client-side computational cost for training a machine-learned model of a given model complexity (e.g., parameter count, computational cost per parameter, etc.) can enable the training of models having increased model complexity for a given client-side computational cost. For example, in some instances, a client-side device (e.g., smartphone, etc.) may have a fixed and limited amount of computational resources (e.g., memory, processor cycles, storage space, etc.) that is available for use in training a machine-learned model. In such instances, reducing a client-side computational cost for a model of a given complexity can enable a larger or more complex model to be trained compared to some alternative methods.

[0057] Additionally, in some instances, increasing a model complexity (e.g., parameter count, etc.) can lead to improved technical performance (e.g., increased accuracy of predictions, increased quality of generative outputs according to a reinforcement learning signal, etc.) due to neural scaling phenomena. For example, in some machine learning model architectures (e.g., neural network architectures, etc.), increased parameter counts can provide increased expressivity (e.g., greater number or variety of functions that can be approximated by the network, etc.) compared to a similar architecture with reduced parameter count. Increased expressivity can in turn lead to improved technical performance (e.g., improved prediction accuracy, etc.) of a machine-learned model (e.g., after training with a sufficiently large quantity of training data, etc.). In this manner, for instance, systems and methods of the present disclosure can in some instances provide improved technical performance of a machine-learned model trained using a given client-side computational cost compared to some alternative training methods.

[0058] Additionally, systems and methods according to example aspects of the present disclosure can provide data security for private, secure, or otherwise access-restricted data on client devices (e.g., increased data security relative to some alternative methods having a similar client-side computational cost, etc.). For example, systems and methods according to some aspects of the present disclosure can train a machine-learned model (e.g., local machine-learned model or component, global or shared machine-learned model, etc.) without sharing private training labels with a server device. Additionally, in some instances, systems and methods according to some aspects of the present disclosure can share one or more update values (e.g., masked update values, aggregated update values, etc.) in a mannerthat secures the data against an adversarial attacker or honest-but-curious server device, while still enabling training of a machine-learned model based on private or secure training labels. In this manner, for instance, systems and methods according to example aspects of the present disclosure can provide increased data security at a given client-side computational cost compared to some alternative methods.

[0059] Additionally, systems and methods according to example aspects of the present disclosure can comprise individual components (e.g., method components, system components, etc.) that may each individually provide a variety of benefits (e.g., technical benefits, practical benefits, etc.) to one or more users or owners of client-side computing devices; one or more users or owners of server-side computing devices; and various other parties. For example, in some instances, a user or owner of a client or server device may benefit directly from a technical benefit (e.g., technical benefit described above, such as reduced computational cost, improved technical performance of a machine-learned model, etc.) or practical benefit provided directly by an example component according to example aspects of the present disclosure. For example, a client device user may receive a direct practical benefit (e.g., longer smartphone battery life, improved user satisfaction, etc.) from any component described herein that may contribute to reduced client-side computational costs of training a machine-learned model, increased data security of private data, and the like. Some client device users may also directly benefit from improved technical performance of a trained machine-learned model, increased size of the machine-learned model, and the like, if the trained machine-learned model is provided (e.g., as a cloud-based service) to some client device users.

[0060] Additionally, in some instances, providing a practical or technical benefit directly to a first person (e.g., user or owner of a client device) may provide an indirect practical or technical benefit to a second person (e.g., user or owner of a server device) interacting with the first person (e.g., in a contractual or business interaction, etc.). As a nonlimiting illustrative example, preventing a server device from exposing private training data may directly benefit a client-side user by protecting the user’s data, and may also benefit a server-side owner by increasing client users’ willingness to provide training data for training a machine-learned model in a privacy -preserving or data-security-preserving way, thereby increasing the server’s access to training data and improving the quality of the trained machine-learned model. In general, it will be understood that providing a benefit to a customer can inherently provide a benefit to a seller and vice versa. For example, providing a benefit to a customer may increase customer satisfaction; increase the customer’s willingnessto return for repeat business; improve the seller’s reputation and therefore its ability to attract new customers; increase a price a customer is willing to pay for a service comprising the benefit; or the like. Similarly, it will be understood that providing a benefit to a seller can inherently provide a benefit to a customer of the seller by reducing a price at which the seller is able or willing to sell; incentivizing the seller to provide improved services, additional features, or improved contractual terms (e.g., non -monetary contractual terms) to the customer; or the like.

[0061] Additionally, in some instances, a first component of a system or method can provide a practical or technical benefit by providing a necessary prerequisite for a second component that provides a practical or technical benefit in combination with the first component. As a non-limiting illustrative example, if a method comprises obtaining data and then using the data to train a machine-learned model, any person or legal entity that benefits from training the machine-learned model inherently benefits from obtaining the data, as obtaining the data was a necessary prerequisite for training the machine-learned model with the data so obtained.

[0062] Example components according to example aspects of the present disclosure that can provide example practical and technical benefits to users and owners of both client and server devices include, but are not limited to, various components and benefits described in this paragraph. In some instances, increased data security of private data of a first computing device, which can directly benefit a first-device user or owner and can indirectly benefit a second-device user or owner, can be provided by each of the following: determining update values on the first device (instead of, e.g., the second device); storing update values locally on the first device; masking or aggregating update values before providing them to the second device; protecting private data by not providing it to the second device; and the like. In some instances, reduced first-device computational cost, which can directly benefit a first- device user by reducing cost and can directly and indirectly benefit a second-device user (e.g., by increasing a maximum model size, etc.), can be provided by each of the following: performing a forward pass using the second device; providing, by the first device to the second device, input data to enable the second device to perform the forward pass; providing a reduced-parameter-count model description to the first computing device (e.g., instead of a full -parameter-count model description), particularly when the reduction in parameter count is large (e.g., 50 percent reduction, 90 percent reduction); providing output activation data from the forward pass to eliminate any need to recompute such data; and the like. In some instances, improved machine-learned model accuracy, which can directly and indirectlybenefit both users and providers of the machine-learned model, can be provided by each of the following: updating a machine-learned model (e.g., on the first device or the second device) based on one or more update values; determining the one or more update values (e.g., based on a gradient of an objective function) to enable updating; providing or obtaining data to enable determining the update values; and the like.

[0063] Various example implementations are described herein with respect to the accompanying Figures.

[0064] Figure l is a block diagram of an example system for training a machine learned model according to example aspects of the present disclosure. The first computing device 102 can provide input data 104 to a second computing system 106 comprising one or more second computing devices. The second computing system 106 can perform a forward pass of the machine-learned model 108 based on the input data 104. The second computing system 106 can provide forward pass data 110 associated with the forward pass to the first computing device 102. Based at least in part on the forward pass data 110, the first computing device 102 can perform a model update determination 112 (e.g., based on private data such as a private or secure training label that is not shared with the second computing system 106) to determine one or more update values for the machine-learned model 108.

[0065] The first computing device 102 can be or include one or more software, firmware, or hardware components configured to perform a model update determination 112 based on forward pass data 110. In some instances, the first computing device 102 can be, comprise, be comprised by, or otherwise share one or more properties with one or more devices described below with respect to Figures 9-17, such as a computing device 99, computing device 98, computing device 50, or other device. In some instances, the first computing device 102 can include a client device, such as a smartphone, laptop, desktop, tablet, smart appliance, or other client device.

[0066] In some instances, input data 104 can be, comprise, be comprised by, or otherwise share one or more properties with inputs 2, described below with respect to Figures 9-10. For example, in some instances, input data 104 can have any property described below with respect to inputs 2. Input data 104 can include data received by the first computing device 102 (e.g., from a user, from another computing device), generated by the first computing device (e.g., using one or more sensors such as cameras or microphones.; using a machine-learned model; etc.), retrieved from computer-readable storage media of the first computing device 102, or obtained by the first computing device 102 in any other manner. In some instances, input data 104 can include unprocessed input data 104 (e.g., unprocessedinputs received from a user), such as unprocessed text, audio, image, video, or other input type. In some instances, input data 104 can include data that has been preprocessed in one or more ways (e.g., for use as input to a machine-learned model 108). As a non-limiting illustrative example, a user may provide input to a first computing device 102 in a natural language format (e.g., English-language text, audio, etc.), and a machine-learned model 108 may be configured to receive vector data associated with a plurality of tokens of the natural language user input. Continuing the non-limiting illustrative example, the first computing device 102 may in some instances provide the user input to the second computing system 106 as input data 104 in a natural language format (e.g., for the second computing system 106 to perform tokenization and other preprocessing). Continuing the non-limiting illustrative example, the first computing device 102 may in other instances perform one or more preprocessing steps (e.g., tokenizing, converting to vector format, etc.) before providing the preprocessed data as input data 104 to the second computing system 106.

[0067] The second computing system 106 can be or include one or more software, firmware, or hardware components configured to generate forward pass data 110 based on input data 104 using a machine-learned model 108. In some instances, the second computing system 106 can be, comprise, be comprised by, or otherwise share one or more properties with one or more devices described below with respect to Figures 9-17, such as a computing device 99, computing device 98, computing device 50, server computing system 60, or other device. In some instances, the second computing system can comprise one computing device or a plurality of computing devices. In some instances, the second computing system 106 can comprise one or more server devices. In some instances, the second computing system 106 can further comprise one or more client devices different from the first computing device 102. For example, some example details of an example implementation in which a second computing system 106 can include one or more client devices are further provided below with respect to Figures 5 A and 5B.

[0068] The machine-learned model 108 can include one or more machine-learned models. The machine-learned model 108 can include various model architectures. An example model architecture for machine-learned model 108 can include a network architecture (e.g., neural network, transformer, Kolmogorov-Arnold network, etc.) comprising one or more nodes (e.g., neurons, etc.) and one or more parameterized edges between the nodes. In some instances, an example model architecture can include one or more layers, each layer comprising one or more nodes. In some instances, a machine-learned model 108 can be or include a differentiable model for which a gradient of a correspondingobjective function or loss function can be computed or estimated. In some instances, a machine-learned model can include one or more layers having sparse neuron activations (e.g., low percentage of non-zero output values for input-layer, hidden-layer, or output-layer neurons) for some values of input data 104. For example, in some instances, a machine- learned model 108 may include a sparse mixture-of-experts model or other model (e.g., sequence processing model such as large language model, etc.) characterized by sparse connections for some or all values of input data 104.

[0069] Performing a forward pass can include, for example, using a machine-learned model 108 to generate one or more machine-learned outputs based on input data 104. For example, a forward pass can include machine-learned inference or prediction; machine- learned generation (e.g., text generation, audio generation, image generation, etc.); machine- learned embedding of input data 104; or any other machine-learned forward pass action. For example, in some instances, a machine-learned model 108 can include a neural network architecture, and a forward pass can include, for each of a plurality of layers, multiplying input activations of the layer by weights of the layer; summing a plurality of weighted inputs at each node of the layer; and applying an activation function (e.g., ReLU, sigmoid, etc.) to the sum at each node to generate an output activation value for the node. A forward pass can include a forward pass of the entire model (e.g., every layer of a plurality of layers of the machine-learned model) or a subset of the model (e.g., a subset of layers of the model, such as every layer but the first or last n layers, etc.).

[0070] Forward pass data 110 can generally include or otherwise represent various types of data. Forward pass data 110 can include data of the same type(s) or of different types of data as compared to input data 104 or other data described herein. In some instances, forward pass data 110 can include various data associated with a forward pass performed by the second computing system 106 based on the input data 104. For example, forward pass data 110 can include a result (e.g., numerical result, binary result, etc.) of any computation (e.g., matrix multiplication computation, activation function computation, multiplication or summation computation, etc.) performed by the second computing system 106 during the forward pass. For example, in some instances, forward pass data 110 can include one or more output activations of one or more nodes (e.g., neurons, etc.) of the machine-learned model 108 during the forward pass. For example, in some instances, forward pass data 110 can include some or all input-layer or hidden-layer output activations of the forward pass having a non-zero impact on one or more output-layer (e.g., final-layer, etc.) output activations of the forward pass.

[0071] In some instances, forward pass data 110 can include a reduced-parametercount model description associated with the forward pass performed by the second computing system 106. A reduced-parameter-count model description can include, for example, a description of a machine-learned model having fewer parameters (e.g., fewer weights, etc.) compared to the full machine-learned model 108. In some instances, the reduced-parametercount model description can describe a machine-learned model that is mathematically equivalent to the full machine-learned model 108 with respect to the forward pass performed by the second computing system 106. For example, the reduced-parameter-count model description can include all parameters (e.g., weights) associated with a non-zero contribution to an output-layer output activation of the forward-pass, and can omit one or more parameters (e.g., weights) having no effect on the output-layer output activation. For example, a reduced- parameter-count model description can omit one or more parameters of the machine-learned model 108 whose omission would have no effect on an output-layer output activation of the forward pass. In some instances, the reduced-parameter-count model description can be configured such that each parameter of the reduced-parameter-count model is a parameter of the machine-learned model 108, and a forward pass of the reduced-parameter-count model based on the input data 104 would be identical to a forward pass of the machine-learned model 108 based on the same input data 104. Additional example details of an example implementation of forward pass data 110, including an example reduced-parameter-count model description, are further provided below with respect to Figure 2.

[0072] A model update determination 112 can include, for example, any method for determining an update value for a machine-learned model based on forward pass data 110. In some instances, a model update determination 112 can include one or more gradient-based update determination methods (e.g., gradient descent, gradient ascent, stochastic gradient descent, Adam optimization, etc.). For example, in some instances, performing a model update determination 112 can include determining a gradient of an objective function (e.g., loss function, etc.) based on the forward pass data 110. In some instances, the forward pass data 110 can include a plurality of weights of the machine-learned model 108 and a plurality of activations that were multiplied by the weights during the forward pass to generate an output value. In some instances, the plurality of weights can be, comprise, or be comprised by a reduced-parameter-count model description of the machine-learned model 108, wherein the reduced-parameter-count model description may be equivalent to the full machine-learned model 108 for the purposes of the training example. In some instances, the first computing device 102 can obtain (e.g., receive from a second computing system 106, retrieve from astorage device of the first computing device, determine, select, etc.) a loss function for training the machine-learned model 108. Based on the forward pass data 110 (e.g., based on a plurality of weights and plurality of activations of the forward pass data 110) and a training label (e.g., secure or private training label), the first computing device 102 can determine (e.g., compute, estimate, stochastically approximate, etc.) a gradient of the objective function for the training example associated with the training label and forward pass data 110. The gradient can include, for example, a plurality of respective partial derivative values associated with a plurality of respective weights of the forward pass data 110. In some instances, determining one or more update values can include determining update values based on the gradient. For example, in some instances, determining an update value can include multiplying the gradient by a learning rate parameter (e.g., adaptive learning rate, etc.). In some instances, determining an update value can include aggregating a gradient with one or more other gradients (e.g., as part of a mini -batch, etc.).

[0073] Figure 2 is a block diagram of an example system for training a machine learned model according to example aspects of the present disclosure. The first computing device 102 can provide input data 104 to a second computing system 106 comprising one or more second computing devices. The second computing system 106 can perform a forward pass of the machine-learned model 108 based on the input data 104. Based on the forward pass, the second computing system 106 can provide a sparse model description 210 and sparse activations 211 to the first computing device 102. Based at least in part on the sparse model description 210 or sparse activations 211, the first computing device 102 can perform a model update determination 112 (e.g., based on a private or secure training label that is not shared with the second computing system 106) to determine one or more update values for the machine-learned model 108.

[0074] A sparse model description 210 can be, comprise, be comprised by, or otherwise share one or more properties with forward pass data 110. For example, in some instances, a sparse model description 210 can be a reduced-parameter-count model description of forward pass data 110. In some instances, the forward pass data 110 may comprise the sparse model description 210 and various other data (e.g., sparse activations 211, etc.). For example, a sparse model description 210 can include a description of a machine-learned model that has fewer parameters than the machine-learned model 108 but is mathematically equivalent to the machine-learned model 108 with respect to a forward pass using input data 104. For example, the sparse model description 210 can include some or all parameters (e.g., weights) associated with a non-zero-valued contribution to an output-layeroutput activation of the forward pass, and can omit one or more parameters (e.g., weights) having no effect (e.g., a zero-valued contribution) on the output-layer output activation. As another example, a reduced-parameter-count model description can omit one or more parameters of the machine-learned model 108 whose omission would have only a small effect (e.g., below a predetermined effect size threshold) on an update value associated with the training example (e.g., small effect on a gradient of an objective function, small effect on an output activation of an output layer, etc.). For example, generating a reduced-parameter-count model description can include identifying all nodes that output an activation value greater than a threshold value (e.g., zero, constant value, threshold hyperparameter that controls a tradeoff between a learning signal and a computational cost, etc.); have an effect size (e.g., size of an effect on another node’s output activation value, effect on one or more gradients or partial derivatives, etc.) greater than a threshold value; or the like.

[0075] As another example, generating a reduced-parameter-count model description can include determining a plurality of parameters based at least in part on a size constraint (e.g., predetermined maximum size, target or preferred size, etc.) such as a parameter count constraint. For example, in some instances, one or more first computing devices 102 (e.g., smartphones, etc.) may be associated with a maximum permissible model size or the like (e.g., based on device constraints, user preferences, etc.). In some instances, generating a reduced-parameter-count model description can include identifying a subset of parameters of the machine-learned model 108 based on a predetermined subset size (e.g., parameter count, percentage of machine-learned model 108 parameters, percentage of parameters of one or more layers, etc.). For example, in some instances, generating a reduced-parameter-count model description can include identifying n nodes of the machine-learned model having the largest output activation values or the largest effect sizes (e.g., size of effect on a final-layer output activation, effect on a gradient or partial derivative, etc.), where n can be a predetermined number or percentage (e.g., 50, 20, 10, 5, 2, or 1 percent, etc.) of parameters of the machine-learned model 108.

[0076] In some instances, a sparse model description 210 can be characterized by sparsity of one or more layers of the described model. For example, in some instances, the sparse model description 210 may describe a reduced-parameter-count model having the same number of layers as the machine-learned model 108 in the same order, but one or more layers (e.g., some or all layers) of the reduced-parameter-count model may have fewer parameters than a corresponding layer of the machine-learned model 108. For example, in some instances, the set of parameters of each of one or more layers of a reduced-parameter-count model may be a strict subset of a corresponding layer of the full machine-learned model 108. For example, in some instances, a machine-learned model 108 may include a model that may have sparse neuron activations (e.g., low percentage of non-zero output values for input-layer, hidden-layer, or output-layer neurons) for some forward passes or some values of input data 104, or for one or more layers of the machine-learned model 108. For example, in some instances, a machine-learned model 108 may include a sparse mixture- of-experts model or other model (e.g., sequence processing model such as large language model, etc.) characterized by sparse connections for some or all values of input data 104. In some instances, one or more layers of the machine-learned model 108 may have a sparsity during the forward pass that is greater than 50 percent; greater than 80 percent; greater than 90 percent; greater than 95 percent; greater than 98 percent; greater than 99 percent; greater than 99.5 percent; greater than 99.9 percent; etc. As used herein a sparsity of a layer during a forward pass can be a percentage of output activations of the layer that are equal to zero during the forward pass. In some instances, one or more layers of the machine-learned model 108 may be configured to have a high (e.g., greater than 50, 80, 90, 95, 98, 99, 99.5, 99.9 percent, etc.) sparsity for all or most values of input data 104, such as one or more layers of a sparse mixture-of-experts model.

[0077] In some instances, a reduced-parameter-count model description can be provided in a variety of formats or data structures. For example, in some instances, a sparse model description 210 can be, comprise, or describe a sparse model graph, such as a graph comprising a plurality of nodes and edges. For example, in some instances, a node of a sparse model graph can correspond to a node of a reduced-parameter-count model (e.g., neuron of a neural network, etc.) and an edge of a sparse model graph can correspond to a parameter (e.g., weight) of the reduced-parameter-count model. A sparse model graph can be stored, for example, in a variety of formats or data structures, such as a triplet-based data structure wherein each edge connecting two nodes is described as a triplet comprising a first node, a second node, and a parameter value (e.g., weight value, etc.) associated with the edge. In some instances, a sparse model graph can be stored or transmitted in a sparse data structure configured to efficiently store or transmit (e.g., using less storage space or less bandwidth) sparse data. For example, a sparse data structure can include a data structure configured to explicitly represent only non-zero values of a sparse data set, with zero values omitted. For example, a data structure may store each non-zero value as a (location, value) pair, wherein a location can include, for example, an index (e.g., vector index, matrix index, tensor index, etc.) associated with the value. As a non-limiting illustrative example, if a machine-learnedmodel 108 comprises 100 layers with 1000 nodes per layer, each edge of a sparse model description 210 description can be represented as a tuplet comprising: a first layer index and first node index describing a location of a first node; a second layer index and second node index describing a location of a second node associated with the edge; and a parameter value (e.g., weight, etc.) associated with the edge. Other data structures are possible.

[0078] Sparse activations 211 can be, comprise, be comprised by, or otherwise share one or more properties with forward pass data 110. For example, in some instances, the forward pass data 110 may comprise the sparse activations 211 and various other data (e.g., sparse model description 210, etc.). In some instances, sparse activations 211 can be stored in a data structure that is the same as or different from a data structure of the sparse model description 210. Continuing the non-limiting illustrative example described above with respect to the sparse model description 210, a tuplet describing an edge of the sparse model description 210 may also comprise an activation value associated with the edge (e.g., an activation value that was multiplied by a weight value of the edge during a forward pass, etc.). In some instances, the sparse activations 211 can include some or all activations associated with a non-zero contribution to an output-layer output activation of the forwardpass, and can omit one or more activations having no effect on the output-layer output activation. For example, the sparse activations 211 can omit one or more activations of the forward pass whose omission would have no effect on an output-layer output activation of the forward pass.

[0079] Figure 3 is a block diagram of an example system for training a machine learned model according to example aspects of the present disclosure. The first computing device 102 can provide input data 104 to a second computing system 106 comprising one or more second computing devices. The second computing system 106 can perform a forward pass of the machine-learned model 108 based on the input data 104. The second computing system 106 can provide forward pass data 110 associated with the forward pass to the first computing device 102. Based at least in part on the forward pass data 110, the first computing device 102 can perform a model update determination 112 (e.g., based on a private or secure training label that is not shared with the second computing system 106) to determine one or more model update values 314 for the machine-learned model 108. The first computing device 102 can provide the model update values 314 to one or more computing devices of the second computing system 106.

[0080] Model update values 314 can include, for example, any value for updating a machine-learned model 108 or any value for determining an update for a machine-learnedmodel 108. For example, in some instances, model update values 314 can include one or more values (e.g., numerical values, quantized values, binary values) for adjusting a corresponding parameter (e.g., weight, etc.) of the machine-learned model 108. Adjusting a parameter can include, for example, adding a corresponding model update value to the parameter. As another example, in some instances, model update values 314 can include one or more values (e.g., raw gradient values, etc.) for determining one or more updates for a machine-learned model 108.

[0081] Providing a model update value 314 to a second computing system 106 can be performed in a variety of ways, including ways that may preserve data security or privacy of one or more private or secure training labels used to determine the model update value 314. In some instances, providing a model update value 314 to a second computing system 106 can include masking (e.g., encrypting, adding noise to, splitting into multiple values, etc.) the model update value 314 before providing it to the second computing system 106. In some instances, providing a model update value 314 to a second computing system 106 can include providing the model update value 314 to a computing device of the second computing system 106 (e.g., second client device, shuffler device, etc.) that is different from a computing device (e.g., server device) that performed the forward pass. Further example details of an example implementation of masked peer-to-peer communication of model update values are provided below with respect to Figures 5A and 5B.

[0082] Figure 4 is a block diagram of an example system for training a machine learned model according to example aspects of the present disclosure. The first computing device 102 can provide input data 104 to a second computing system 106 comprising one or more second computing devices. The second computing system 106 can perform a forward pass of the machine-learned model 108 based on the input data 104. The second computing system 106 can provide forward pass data 110 associated with the forward pass to the first computing device 102. Based at least in part on the forward pass data 110, the first computing device 102 can perform a model update determination 112 (e.g., based on a private or secure training label that is not shared with the second computing system 106) to determine one or more model update values 314 for the machine-learned model 108. Based on the model update values 314, the first computing device can update a local machine-learned model 408 stored locally on the first computing device 102.

[0083] The machine-learned model 408 can be, comprise, be comprised by, share one or more properties with, or otherwise be associated with the machine-learned model 108. For example, in some instances, the machine-learned model 408 can include a local copy of all orpart of the machine-learned model 108. As another example, in some instances, the machine- learned model 408 can include one or more adapters (e.g., rank decomposition matrices for low-rank adaptation, etc.) for adjusting one or more parameters of the machine-learned model 108. As another example, in some instances, a machine-learned model 108 and machine- learned model 408 can comprise separate components of a combined model. As a nonlimiting illustrative example, a machine-learned model 108 can comprise one or more first layers of a combined machine-learned model. In some instances, the first layers can be “global” layers that are shared between a plurality of first computing devices 102 and the second computing system 106. Continuing the non-limiting illustrative example, a machine- learned model 408 can comprise one or more additional layers of the combined machine- learned model. In some instances, the additional layers can be “local” layers comprising private parameter values that may be exclusive to a single first computing device 102.

[0084] Performing inference with a combined model comprising a global component and local component can include, for example, performing a first partial forward pass using the machine-learned model 108; providing an output of the first partial forward pass to the machine-learned model 408; and performing a second partial forward pass using the machine-learned model 408 based on the output of the first forward pass. In such instances, the first partial forward pass can be performed by the second computing system 106, and the forward pass data 110 may represent data associated with the first partial forward pass. In some instances, the first computing device 102 can perform the second partial forward pass based on the forward pass data 110. In some instances, the first computing device 102 can perform a model update determination 112 based on the forward pass data 110 and data from the second partial forward pass. Updating a machine-learned model 408 can include, for example, adding each model update value 314 to a corresponding parameter of the machine- learned model 408, or otherwise adjusting one or more parameters of the machine-learned model 408 based on the model update values 314.

[0085] Figure 5A is a first block diagram of an example system for communicating model update values according to example aspects of the present disclosure. A plurality of client computing devices 502 can each provide input data 104 to a server computing system 506 comprising one or more server computing devices. The server computing system 506 can perform a forward pass of the machine-learned model 108 based on each set of input data 104. The server computing system 506 can provide forward pass data 110 associated with the forward pass to the client computing devices 502. Based at least in part on the forward pass data 110, each client computing device 502 can perform a model update determination 112(e.g., based on a private or secure training label that is not shared with the server computing system 506) to determine one or more model update values 314 for the machine-learned model 108.

[0086] In some instances, a client device 502 can be, comprise, be comprised by, or otherwise share one or more properties with a first device 102. In some instances, a client device 502 can be, comprise, be comprised by, or otherwise share one or more properties with a device of a second computing system 106. For example, in some instances, any one or more of the client devices 502a-z can act as a first computing device 102 when providing input data 104 to the server computing system 506 and receiving forward pass data 110 from the server computing system 506. Additionally, any one or more of the client devices 502a-z can act as a device of a second computing system 106 when receiving model update values 314, 514 from another client device 502 (e.g., as depicted below with respect to Figure 5B).

[0087] In some instances, a server computing system 506 can be, comprise, be comprised by, or otherwise share one or more properties with a second computing system 106. For example, in some instances, a server computing system 506 can be, comprise, be comprised by, or otherwise share one or more properties with one or more devices described below with respect to Figures 9-17, such as a computing device 99, computing device 98, computing device 50, server computing system 60, or other device. In some instances, a server computing system 506 can comprise one or more server devices.

[0088] Figure 5B is a second block diagram of an example system for communicating model update values according to example aspects of the present disclosure. Subsequent to determining one or more model update values 314 (e.g., as depicted in FIG. 5A), each client computing device 502 can provide one or more model update values 314 or masked model update values 514 to one or more other client computing devices 502. In some instances, one or more of the client computing devices 502 can aggregate (e.g., sum, combine, etc.) the model update values 314 or masked update values 514 to generate one or more aggregated update values 516. In some instances, one or more of the client computing devices 502 can provide the aggregated update values 516 to the server computing system 506. In some instances, the server computing system 506 can update the machine-learned model 108 based on the aggregated update values 516.

[0089] Masked model update values 514 can include, for example, any data determined based at least in part on model update values 314. In some instances, masked model update values 514 can include model update values 314 that have been masked (e.g., encrypted, etc.) according to a security protocol, such that individual model update values314 cannot be learned by an honest-but-curious device or an adversarial attacker among the second computing system 106, server computing system 506, or other client computing devices 502. In some instances, masked model update values 514 can include model update values 314 that have been masked (e.g., noised, etc.) according to a differential privacy protocol, such that a machine-learned model 108 can learn aggregate values associated with a plurality of client devices 502 while limiting information that can be learned by the machine- learned model 108 about any individual client device 502. Masking can include a variety of activities for masking, disguising, or otherwise preventing direct access to the model update values 314, including but not limited to adding random noise to the model update values 314; encrypting the model update values 314 (e.g., according to a secure aggregation protocol); splitting the model update values 314 into a plurality of component values; aggregating the model update values 314 with a plurality of other model update values; or the like. As a nonlimiting illustrative example, masking can include adding sufficient random noise to turn sparse model update values 314 (e.g., associated with a sparse model description 210) into a dense model update (e.g., comprising non-zero parameter update values for all or most parameters of the machine-learned model 108). As another non-limiting illustrative example, a differential privacy protocol can include generating a plurality of n random masked model update values 514 (e.g., dense model updates, etc.), along with a final masked model update value 514 configured to cause the sum of the (n + 1) masked model update values 514 to equal the model update values 314; and sharing each of the (n + 1) masked model update values 514 with one of (n + 1) other client devices 502; wherein each client device aggregates the masked model update values 514 they receive to generate an aggregated model update value 516.

[0090] Aggregated model update values 516 can include, for example, an aggregation of a plurality of model update values 314 or masked model update values 514. An aggregation can include, for example, a sum, product, cryptographic combination, or other combination of a plurality of values. In some instances, an aggregated model update value 516 can be a masked value (e.g., noised value, etc.) configured to be combined with other aggregated model update values 516 to generate a batch model update based on a plurality of model update values 314. For example, in some instances, a secure aggregation protocol or differential privacy protocol can be configured such that a batch model update based on a plurality of aggregated model update values 516 is mathematically equivalent or substantially similar to (e.g., within a specified error margin, etc.) updating based on the true model update values 314. In some instances, a batch size of a batch model update can be configured toprevent an honest-but-curious or adversarial server computing system 506 from inferring private data of individual client devices 502 (e.g., private training labels used to determine the model update values 314). For example, in some instances, a minimum batch size can be determined based at least in part on a sparsity of the machine-learned model 108. For example, in some instances, a minimum batch size can be determined such that each updated parameter of the machine-learned model 108 depends on a plurality of contributing model update values 314 from a plurality of contributing first computing devices 102, such that no individual training label (e.g., private or secure training label) is likely to be inferred (e.g., according to a confidence threshold, etc.) from any batch update value.Example Methods

[0091] Figure 6 depicts a flowchart diagram of an example method for training a machine-learned model according to example embodiments of the present disclosure. Although Figure 6 depicts steps performed in a particular order for purposes of illustration and discussion, the methods of the present disclosure are not limited to the particularly illustrated order or arrangement. The various steps of example method 600 can be omitted, rearranged, combined, and / or adapted in various ways without deviating from the scope of the present disclosure.

[0092] At 602, example method 600 can include providing, by a first computing device (e.g., first computing device 102) to a computing system comprising one or more second computing devices (e.g., second computing system 106), data (e.g., input data 104) indicative of one or more input values for one or more layers of a machine-learned model (e.g., machine-learned model 108). In some instances, example method 600 at 602 can include using one or more systems or performing one or more activities described with respect to Figures 1-5B.

[0093] At 604, example method 600 can include receiving, by the first computing device from the computing system, data (e.g., forward pass data 110) indicative of a forward pass of the machine-learned model based on the one or more input values, wherein the data indicative of the forward pass comprises a reduced-parameter-count model description (e.g., sparse model description 210, etc.) associated with the machine-learned model. In some instances, example method 600 at 604 can include using one or more systems or performing one or more activities described with respect to Figures 1-5B.

[0094] At 606, example method 600 can include determining, by the first computing device based on the data indicative of the forward pass, one or more update values (e.g.,model update values 314, masked update values 514, aggregated update values 516, etc.) for the machine-learned model, wherein the one or more update values comprise at least one update value for at least one parameter of the reduced-parameter-count model description. In some instances, example method 600 at 606 can include using one or more systems or performing one or more activities described with respect to Figures 1-5B.

[0095] At 608a, example method 600 can include providing, by the first computing device, the one or more update values to the computing system. In some instances, example method 600 at 608a can include using one or more systems or performing one or more activities described with respect to Figures 1-5B.

[0096] At 608b, example method 600 can include updating, by the first computing device based on the one or more update values, one or more local parameters (e.g., machine- learned model 408) stored locally on the first computing device, wherein the local parameters are associated with the machine-learned model. In some instances, example method 600 at 608b can include using one or more systems or performing one or more activities described with respect to Figures 1-5B.

[0097] Figure 7 depicts a flowchart diagram of an example method for training a machine-learned model according to example embodiments of the present disclosure. Although Figure 7 depicts steps performed in a particular order for purposes of illustration and discussion, the methods of the present disclosure are not limited to the particularly illustrated order or arrangement. The various steps of example method 700 can be omitted, rearranged, combined, and / or adapted in various ways without deviating from the scope of the present disclosure.

[0098] At 702, example method 700 can include receiving, from a first computing device (e.g., first computing device 102), data (e.g., input data 104) indicative of one or more input values for one or more layers of a machine-learned model (e.g., machine-learned model 108). In some instances, example method 700 at 702 can include using one or more systems or performing one or more activities described with respect to Figures 1-5B.

[0099] At 704, example method 700 can include performing a forward pass of the machine-learned model based on the one or more input values. In some instances, example method 700 at 704 can include using one or more systems or performing one or more activities described with respect to Figures 1-5B.

[0100] At 706, example method 700 can include providing (e.g., by a second computing system 106), to the first computing device, data (e.g., forward pass data 110) indicative of the forward pass, wherein the data indicative of the forward pass comprises areduced-parameter-count model description (e.g., sparse model description 210, etc.) of the machine-learned model. In some instances, example method 700 at 706 can include using one or more systems or performing one or more activities described with respect to Figures 1-5B.

[0101] At 708, example method 700 can include receiving, from the first computing device, data indicative of one or more update values (e.g., model update values 314, aggregated update values 516, etc.) for the machine-learned model. In some instances, example method 700 at 708 can include using one or more systems or performing one or more activities described with respect to Figures 3, 5 A, and 5B.

[0102] At 710, example method 700 can include updating the machine-learned model based at least in part on the data indicative of the update values. In some instances, example method 700 at 710 can include using one or more systems or performing one or more activities described with respect to Figures 3, 5 A, and 5B.

[0103] Figure 8 depicts a flowchart of a method 800 for training one or more machine-learned models according to aspects of the present disclosure. For instance, an example machine-learned model can include a machine-learned model 108 or machine- learned model 408.

[0104] One or more portion(s) of example method 800 can be implemented by a computing system that includes one or more computing devices such as, for example, computing systems described with reference to the other figures. Each respective portion of example method 800 can be performed by any (or any combination) of one or more computing devices. Moreover, one or more portion(s) of example method 800 can be implemented on the hardware components of the device(s) described herein, for example, to train one or more systems or models. Figure 8 depicts elements performed in a particular order for purposes of illustration and discussion. Those of ordinary skill in the art, using the disclosures provided herein, will understand that the elements of any of the methods discussed herein can be adapted, rearranged, expanded, omitted, combined, or modified in various ways without deviating from the scope of the present disclosure. Figure 8 is described with reference to elements / terms described with respect to other systems and figures for exemplary illustrated purposes and is not meant to be limiting. One or more portions of example method 800 can be performed additionally, or alternatively, by other systems.

[0105] At 802, example method 800 can include obtaining a training instance. A set of training data can include a plurality of training instances divided between multiple datasets (e.g., a training dataset, a validation dataset, or testing dataset). A training instance can be labeled or unlabeled. Although referred to in example method 800 as a “training” instance, itis to be understood that runtime inferences can form training instances when a model is trained using an evaluation of the model’s performance on that runtime instance (e.g., online training / learning). Example data types for the training instance and various tasks associated therewith are described throughout the present disclosure.

[0106] At 804, example method 800 can include processing, using one or more machine-learned models, the training instance to generate an output. The output can be directly obtained from the one or more machine-learned models or can be a downstream result of a chain of processing operations that includes an output of the one or more machine- learned models.

[0107] At 806, example method 800 can include receiving an evaluation signal associated with the output. The evaluation signal can be obtained using a loss function. Various determinations of loss can be used, such as mean squared error, likelihood loss, cross entropy loss, hinge loss, contrastive loss, or various other loss functions. The evaluation signal can be computed using known ground-truth labels (e.g., supervised learning), predicted or estimated labels (e.g., semi- or self-supervised learning), or without labels (e.g., unsupervised learning). The evaluation signal can be a reward (e.g., for reinforcement learning). The reward can be computed using a machine-learned reward model configured to generate rewards based on output(s) received. The reward can be computed using feedback data describing human feedback on the output(s).

[0108] At 808, example method 800 can include updating the machine-learned model using the evaluation signal. For example, values for parameters of the machine-learned model(s) can be learned, in some embodiments, using various training or learning techniques, such as, for example, backwards propagation. For example, the evaluation signal can be backpropagated from the output (or another source of the evaluation signal) through the machine-learned model(s) to update one or more parameters of the model(s) (e.g., based on a gradient of the evaluation signal with respect to the parameter value(s)). For example, system(s) containing one or more machine-learned models can be trained in an end-to-end manner. Gradient descent techniques can be used to iteratively update the parameters over a number of training iterations. In some implementations, performing backwards propagation of errors can include performing truncated backpropagation through time. Example method 800 can include implementing a number of generalization techniques (e.g., weight decays, dropouts, etc.) to improve the generalization capability of the models being trained.

[0109] In some implementations, example method 800 can be implemented for training a machine-learned model from an initialized state to a fully trained state (e.g., whenthe model exhibits a desired performance profile, such as based on accuracy, precision, recall, etc.).

[0110] In some implementations, example method 800 can be implemented for particular stages of a training procedure. For instance, in some implementations, example method 800 can be implemented for pre-training a machine-learned model. Pre-training can include, for instance, large-scale training over potentially noisy data to achieve a broad base of performance levels across a variety of tasks / data types. In some implementations, example method 800 can be implemented for fine-tuning a machine-learned model. Fine-tuning can include, for instance, smaller-scale training on higher-quality (e.g., labeled, curated, etc.) data. Fine-tuning can affect all or a portion of the parameters of a machine-learned model. For example, various portions of the machine-learned model can be “frozen” for certain training stages. For example, parameters associated with an embedding space can be “frozen” during fine-tuning (e.g., to retain information learned from a broader domain(s) than present in the fine-tuning dataset(s)). An example fine-tuning approach includes reinforcement learning. Reinforcement learning can be based on user feedback on model performance during use.Example Machine-Learned Models

[0111] Figure 9 is a block diagram of an example processing flow for using machine- learned model(s) 1 to process input(s) 2 to generate output(s) 3.

[0112] Machine-learned model(s) 1 can be or include one or multiple machine- learned models or model components. Example machine-learned models can include neural networks (e.g., deep neural networks). Example machine-learned models can include nonlinear models or linear models. Example machine-learned models can use other architectures in lieu of or in addition to neural networks. Example machine-learned models can include decision tree based models, support vector machines, hidden Markov models, Bayesian networks, linear regression models, k-means clustering models, etc.

[0113] Example neural networks can include feed-forward neural networks, recurrent neural networks (RNNs), including long short-term memory (LSTM) based recurrent neural networks, convolutional neural networks (CNNs), diffusion models, generative-adversarial networks, or other forms of neural networks. Example neural networks can be deep neural networks. Some example machine-learned models can leverage an attention mechanism such as self-attention. For example, some example machine-learned models can include multiheaded self-attention models.

[0114] Machine-learned model(s) 1 can include a single or multiple instances of the same model configured to operate on data from input(s) 2. Machine-learned model(s) 1 can include an ensemble of different models that can cooperatively interact to process data from input(s) 2. For example, machine-learned model(s) 1 can employ a mixture-of-experts structure. See, e.g., Zhou et al., Mixture-of-Experts with Expert Choice Routing, ARXIV:2202.09368V2 (Oct. 14, 2022).

[0115] Input(s) 2 can generally include or otherwise represent various types of data. Input(s) 2 can include one type or many different types of data. Output(s) 3 can be data of the same type(s) or of different types of data as compared to input(s) 2. Output(s) 3 can include one type or many different types of data.

[0116] Example data types for input(s) 2 or output(s) 3 include natural language text data, software code data (e.g., source code, object code, machine code, or any other form of computer-readable instructions or programming languages), machine code data (e.g., binary code, assembly code, or other forms of machine-readable instructions that can be executed directly by a computer's central processing unit), assembly code data (e.g., low-level programming languages that use symbolic representations of machine code instructions to program a processing unit), genetic data or other chemical or biochemical data, image data, audio data, audiovisual data, haptic data, biometric data, medical data, financial data, statistical data, geographical data, astronomical data, historical data, sensor data generally (e.g., digital or analog values, such as voltage or other absolute or relative level measurement values from a real or artificial input, such as from an audio sensor, light sensor, displacement sensor, etc.), and the like. Data can be raw or processed and can be in any format or schema.

[0117] In multimodal inputs 2 or outputs 3, example combinations of data types include image data and audio data, image data and natural language data, natural language data and software code data, image data and biometric data, sensor data and medical data, etc. It is to be understood that any combination of data types in an input 2 or an output 3 can be present.

[0118] An example input 2 can include one or multiple data types, such as the example data types noted above. An example output 3 can include one or multiple data types, such as the example data types noted above. The data type(s) of input 2 can be the same as or different from the data type(s) of output 3. It is to be understood that the example data types noted above are provided for illustrative purposes only. Data types contemplated within the scope of the present disclosure are not limited to those examples noted above.Example Machine-Learned Sequence Processing Models

[0119] Figure 10 is a block diagram of an example implementation of an example machine-learned model configured to process sequences of information. For instance, an example implementation of machine-learned model(s) 1 can include machine-learned sequence processing model(s) 4. An example system can pass input(s) 2 to sequence processing model(s) 4. Sequence processing model(s) 4 can include one or more machine- learned components. Sequence processing model(s) 4 can process the data from input(s) 2 to obtain an input sequence 5. Input sequence 5 can include one or more input elements 5-1, 5- 2, . . . , 5-A , etc. obtained from input(s) 2. Sequence processing model 4 can process input sequence 5 using prediction layer(s) 6 to generate an output sequence 7. Output sequence 7 can include one or more output elements 7-1, 7-2, . . . , 7-A, etc. generated based on input sequence 5. The system can generate output(s) 3 based on output sequence 7.

[0120] Sequence processing model(s) 4 can include one or multiple machine-learned model components configured to ingest, generate, or otherwise reason over sequences of information. For example, some example sequence processing models in the text domain are referred to as “Large Language Models,” or LLMs. See, e.g., PaLM 2 Technical Report, GOOGLE, https: / / ai.google / static / documents / palm2techreport.pdf (n.d.). Other example sequence processing models can operate in other domains, such as image domains, see, e.g., Dosovitskiy et al., An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, ARXIV:2010.11929v2 (Jun. 3, 2021), audio domains, see, e.g., Agostinelli et al., MusicLM: Generating Music From Text, ARXIV:2301.1 1325V1 (Jan. 26, 2023), biochemical domains, see, e.g., Jumper et al., Highly accurate protein structure prediction with AlphaFold, 596 Nature 583 (Aug. 26, 2021), by way of example. Sequence processing model(s) 4 can process one or multiple types of data simultaneously. Sequence processing model(s) 4 can include relatively large models (e.g., more parameters, computationally expensive, etc.), relatively small models (e.g., fewer parameters, computationally lightweight, etc.), or both.

[0121] In general, sequence processing model(s) 4 can obtain input sequence 5 using data from input(s) 2. For instance, input sequence 5 can include a representation of data from input(s) 2 in a format understood by sequence processing model(s) 4. One or more machine- learned components of sequence processing model(s) 4 can ingest the data from input(s) 2, parse the data into pieces compatible with the processing architectures of sequence processing model(s) 4 (e.g., via “tokenization”), and project the pieces into an input space associated with prediction layer(s) 6 (e.g., via “embedding”).

[0122] Sequence processing model(s) 4 can ingest the data from input(s) 2 and parse the data into a sequence of elements to obtain input sequence 5. For example, a portion of input data from input(s) 2 can be broken down into pieces that collectively represent the content of the portion of the input data. The pieces can provide the elements of the sequence.

[0123] Elements 5-1, 5-2, . . . , 5-M can represent, in some cases, building blocks for capturing or expressing meaningful information in a particular data domain. For instance, the elements can describe “atomic units” across one or more domains. For example, for textual input source(s), the elements can correspond to groups of one or more words or sub-word components, such as sets of one or more characters.

[0124] For example, elements 5-1, 5-2, . . . , 5-M can represent tokens obtained using a tokenizer. For instance, a tokenizer can process a given portion of an input source and output a series of tokens (e.g., corresponding to input elements 5-1, 5-2, . . . , 5-A ) that represent the portion of the input source. Various approaches to tokenization can be used. For instance, textual input source(s) can be tokenized using a byte-pair encoding (BPE) technique. See, e.g., Kudo et al., SentencePiece: A simple and language independent subword tokenizer and detokenizer for Neural Text Processing, PROCEEDINGS OF THE 2018 CONFERENCE ON EMPIRICAL METHODS IN NATURAL LANGUAGE PROCESSING (System Demonstrations), pages 66-71 (October 31-November 4, 2018), https: / / aclanthology.org / D18-2012.pdf. Image-based input source(s) can be tokenized by extracting and serializing patches from an image.

[0125] In general, arbitrary data types can be serialized and processed into input sequence 5. It is to be understood that element(s) 5-1, 5-2, . . . , 5-M depicted in Figure 10 can be the tokens or can be the embedded representations thereof.

[0126] Prediction layer(s) 6 can predict one or more output elements 7-1, 7-2, . . . , 7- N based on the input elements. Prediction layer(s) 6 can include one or more machine-learned model architectures, such as one or more layers of learned parameters that manipulate and transform the input(s) to extract higher-order meaning from, and relationships between, input element(s) 5-1, 5-2, . . . , 5-M. In this manner, for instance, example prediction layer(s) 6 can predict new output element(s) in view of the context provided by input sequence 5.

[0127] Prediction layer(s) 6 can evaluate associations between portions of input sequence 5 and a particular output element. These associations can inform a prediction of the likelihood that a particular output follows the input context. For example, consider the textual snippet, “The carpenter’s toolbox was small and heavy. It was full of .” Example prediction layer(s) 6 can identify that “It” refers back to “toolbox” by determining arelationship between the respective embeddings. Example prediction layer(s) 6 can also link “It” to the attributes of the toolbox, such as “small” and “heavy.” Based on these associations, prediction layer(s) 6 can, for instance, assign a higher probability to the word “nails” than to the word “sawdust.”

[0128] A transformer is an example architecture that can be used in prediction layer(s) 4. See, e.g., Vaswani et al., Atention Is All You Need, ARXIV: 1706.03762v7 (Aug. 2, 2023). A transformer is an example of a machine-learned model architecture that uses an attention mechanism to compute associations between items within a context window. The context window can include a sequence that contains input sequence 5 and potentially one or more output element(s) 7-1, 7-2, . . . , 7-N. A transformer block can include one or more attention layer(s) and one or more post-attention layer(s) (e.g., feedforward layer(s), such as a multi-layer perceptron).

[0129] Prediction layer(s) 6 can include other machine-learned model architectures in addition to or in lieu of transformer-based architectures. For example, recurrent neural networks (RNNs) and long short-term memory (LSTM) models can also be used, as well as convolutional neural networks (CNNs). In general, prediction layer(s) 6 can leverage various kinds of artificial neural networks that can understand or generate sequences of information.

[0130] Output sequence 7 can include or otherwise represent the same or different data types as input sequence 5. For instance, input sequence 5 can represent textual data, and output sequence 7 can represent textual data. Input sequence 5 can represent image, audio, or audiovisual data, and output sequence 7 can represent textual data (e.g., describing the image, audio, or audiovisual data). It is to be understood that prediction layer(s) 6, and any other interstitial model components of sequence processing model(s) 4, can be configured to receive a variety of data types in input sequence(s) 5 and output a variety of data types in output sequence(s) 7.

[0131] Output sequence 7 can have various relationships to input sequence 5. Output sequence 7 can be a continuation of input sequence 5. Output sequence 7 can be complementary to input sequence 5. Output sequence 7 can translate, transform, augment, or otherwise modify input sequence 5. Output sequence 7 can answer, evaluate, confirm, or otherwise respond to input sequence 5. Output sequence 7 can implement (or describe instructions for implementing) an instruction provided via input sequence 5.

[0132] Output sequence 7 can be generated autoregressively. For instance, for some applications, an output of one or more prediction layer(s) 6 can be passed through one or more output layers (e.g., softmax layer) to obtain a probability distribution over an outputvocabulary (e.g., a textual or symbolic vocabulary) conditioned on a set of input elements in a context window. In this manner, for instance, output sequence 7 can be autoregressively generated by sampling a likely next output element, adding that element to the context window, and re-generating the probability distribution based on the updated context window, and sampling a likely next output element, and so forth.

[0133] Output sequence 7 can also be generated non-autoregressively. For instance, multiple output elements of output sequence 7 can be predicted together without explicit sequential conditioning on each other. See, e.g., Saharia et al., Non-Autoregressive Machine Translation with Latent Alignments, ARXIV:2004.07437V3 (NOV. 16, 2020).

[0134] Output sequence 7 can include one or multiple portions or elements. In an example content generation configuration, output sequence 7 can include multiple elements corresponding to multiple portions of a generated output sequence (e.g., a textual sentence, values of a discretized waveform, computer code, etc.). In an example classification configuration, output sequence 7 can include a single element associated with a classification output. For instance, an output “vocabulary” can include a set of classes into which an input sequence is to be classified. For instance, a vision transformer block can pass latent state information to a multilayer perceptron that outputs a likely class value associated with an input image.

[0135] Figure 11 is a block diagram of an example technique for populating an example input sequence 8. Input sequence 8 can include various functional elements that form part of the model infrastructure, such as an element 8-0 obtained from a task indicator 9 that signals to any model(s) that process input sequence 8 that a particular task is being performed (e.g., to help adapt a performance of the model(s) to that particular task). Input sequence 8 can include various data elements from different data modalities. For instance, an input modality 10-1 can include one modality of data. A data-to- sequence model 11-1 can process data from input modality 10-1 to project the data into a format compatible with input sequence 8 (e.g., one or more vectors dimensioned according to the dimensions of input sequence 8) to obtain elements 8-1, 8-2, 8-3. Another input modality 10-2 can include a different modality of data. A data-to-sequence model 11-2 can project data from input modality 10-2 into a format compatible with input sequence 8 to obtain elements 8-4, 8-5, 8- 6. Another input modality 10-3 can include yet another different modality of data. A data-to- sequence model 11-3 can project data from input modality 10-3 into a format compatible with input sequence 8 to obtain elements 8-7, 8-8, 8-9.

[0136] Input sequence 8 can be the same as or different from input sequence 5. Input sequence 8 can be a multimodal input sequence that contains elements that represent data from different modalities using a common dimensional representation. For instance, an embedding space can have P dimensions. Input sequence 8 can be configured to contain a plurality of elements that have / Jdimensions. In this manner, for instance, example implementations can facilitate information extraction and reasoning across diverse data modalities by projecting data into elements in the same embedding space for comparison, combination, or other computations therebetween.

[0137] For example, elements 8-0, . . . , 8-9 can indicate particular locations within a multidimensional embedding space. Some elements can map to a set of discrete locations in the embedding space. For instance, elements that correspond to discrete members of a predetermined vocabulary of tokens can map to discrete locations in the embedding space that are associated with those tokens. Other elements can be continuously distributed across the embedding space. For instance, some data types can be broken down into continuously defined portions (e.g., image patches) that can be described using continuously distributed locations within the embedding space.

[0138] In some implementations, the expressive power of the embedding space may not be limited to meanings associated with any particular set of tokens or other building blocks. For example, a continuous embedding space can encode a spectrum of high-order information. An individual piece of information (e.g., a token) can map to a particular point in that space: for instance, a token for the word “dog” can be projected to an embedded value that points to a particular location in the embedding space associated with canine-related information. Similarly, an image patch of an image of a dog on grass can also be projected into the embedding space. In some implementations, the projection of the image of the dog can be similar to the projection of the word “dog” while also having similarity to a projection of the word “grass,” while potentially being different from both. In some implementations, the projection of the image patch may not exactly align with any single projection of a single word. In some implementations, the projection of the image patch can align with a combination of the projections of the words “dog” and “grass.” In this manner, for instance, a high-order embedding space can encode information that can be independent of data modalities in which the information is expressed.

[0139] Task indicator 9 can include a model or model component configured to identify a task being performed and inject, into input sequence 8, an input value represented by element 8-0 that signals which task is being performed. For instance, the input value canbe provided as a data type associated with an input modality and projected along with that input modality (e.g., the input value can be a textual task label that is embedded along with other textual data in the input; the input value can be a pixel-based representation of a task that is embedded along with other image data in the input; etc.). The input value can be provided as a data type that differs from or is at least independent from other input(s). For instance, the input value represented by element 8-0 can be learned within a continuous embedding space.

[0140] Input modalities 10-1, 10-2, and 10-3 can be associated with various different data types (e.g., as described above with respect to input(s) 2 and output(s) 3).

[0141] Data-to-sequence models 11-1, 11-2, and 11-3 can be the same or different from each other. Data-to-sequence models 11-1, 11-2, and 11-3 can be adapted to each respective input modality 10-1, 10-2, and 10-3. For example, a textual data-to-sequence model can subdivide a portion of input text and project the subdivisions into element(s) in input sequence 8 (e.g., elements 8-1, 8-2, 8-3, etc.). An image data-to-sequence model can subdivide an input image and project the subdivisions into element(s) in input sequence 8 (e.g., elements 8-4, 8-5, 8-6, etc.). An arbitrary datatype data-to-sequence model can subdivide an input of that arbitrary datatype and project the subdivisions into element(s) in input sequence 8 (e.g., elements 8-7, 8-8, 8-9, etc.).

[0142] Data-to-sequence models 11-1, 11-2, and 11-3 can form part of machine- learned sequence processing model(s) 4. Data-to-sequence models 11-1, 11-2, and 11-3 can be jointly trained with or trained independently from machine-learned sequence processing model(s) 4. Data-to-sequence models 11-1, 11-2, and 11-3 can be trained end-to-end with machine-learned sequence processing model(s) 4.Example Machine-Learned Model Development Platform

[0143] Figure 12 is a block diagram of an example model development platform 12 that can facilitate creation, adaptation, and refinement of example machine-learned models (e.g., machine-learned model(s) 1, sequence processing model(s) 4, etc.). Model development platform 12 can provide a number of different toolkits that developer systems can employ in the development of new or adapted machine-learned models.

[0144] Model development platform 12 can provide one or more model libraries 13 containing building blocks for new models. Model libraries 13 can include one or more pretrained foundational models 13-1, which can provide a backbone of processing power across various tasks. Model libraries 13 can include one or more pre-trained expert models 13-2,which can be focused on performance in particular domains of expertise. Model libraries 13 can include various model primitives 13-3, which can provide low-level architectures or components (optionally pre-trained), which can be assembled in various arrangements as desired.

[0145] Model development platform 12 can receive selections of various model components 14. Model development platform 12 can pass selected model components 14 to a workbench 15 that combines selected model components 14 into a development model 16.

[0146] Workbench 15 can facilitate further refinement and adaptation of development model 16 by leveraging a number of different toolkits integrated with model development platform 12. For example, workbench 15 can facilitate alignment of the development model 16 with a desired performance profile on various tasks using a model alignment toolkit 17.

[0147] Model alignment toolkit 17 can provide a number of tools for causing development model 16 to generate outputs aligned with desired behavioral characteristics. Alignment can include increasing an accuracy, precision, recall, etc. of model outputs. Alignment can include enforcing output styles, schema, or other preferential characteristics of model outputs. Alignment can be general or domain-specific. For instance, a pre-trained foundational model 13-1 can begin with an initial level of performance across multiple domains. Alignment of the pre-trained foundational model 13-1 can include improving a performance in a particular domain of information or tasks (e.g., even at the expense of performance in another domain of information or tasks).

[0148] Model alignment toolkit 17 can integrate one or more dataset(s) 17-1 for aligning development model 16. Curated dataset(s) 17-1 can include labeled or unlabeled training data. Dataset(s) 17-1 can be obtained from public domain datasets. Dataset(s) 17-1 can be obtained from private datasets associated with one or more developer system(s) for the alignment of bespoke machine-learned model(s) customized for private use-cases.

[0149] Pre-training pipelines 17-2 can include a machine-learned model training workflow configured to update development model 16 over large-scale, potentially noisy datasets. For example, pre-training can leverage unsupervised learning techniques (e.g., denoising, etc.) to process large numbers of training instances to update model parameters from an initialized state and achieve a desired baseline performance. Pre-training pipelines 17-2 can leverage unlabeled datasets in dataset(s) 17-1 to perform pre-training. Workbench 15 can implement a pre-training pipeline 17-2 to pre-train development model 16.

[0150] Fine-tuning pipelines 17-3 can include a machine-learned model training workflow configured to refine the model parameters of development model 16 with higher-quality data. Fine-tuning pipelines 17-3 can update development model 16 by conducting supervised training with labeled dataset(s) in dataset(s) 17-1. Fine-tuning pipelines 17-3 can update development model 16 by conducting reinforcement learning using reward signals from user feedback signals. Workbench 15 can implement a fine-tuning pipeline 17-3 to finetune development model 16.

[0151] Prompt libraries 17-4 can include sets of inputs configured to induce behavior aligned with desired performance criteria. Prompt libraries 17-4 can include few-shot prompts (e.g., inputs providing examples of desired model outputs for prepending to a desired runtime query), chain-of-thought prompts (e.g., inputs providing step-by-step reasoning within the exemplars to facilitate thorough reasoning by the model), and the like.

[0152] Example prompts can be retrieved from an available repository of prompt libraries 17-4. Example prompts can be contributed by one or more developer systems using workbench 15.

[0153] In some implementations, pre-trained or fine-tuned models can achieve satisfactory performance without exemplars in the inputs. For instance, zero-shot prompts can include inputs that lack exemplars. Zero-shot prompts can be within a domain within a training dataset or outside of the training domain(s).

[0154] Prompt libraries 17-4 can include one or more prompt engineering tools.Prompt engineering tools can provide workflows for retrieving or learning optimized prompt values. Prompt engineering tools can facilitate directly learning prompt values (e.g., input element values) based on one or more training iterations. Workbench 15 can implement prompt engineering tools in development model 16.

[0155] Prompt libraries 17-4 can include pipelines for prompt generation. For example, inputs can be generated using development model 16 itself or other machine- learned models. In this manner, for instance, a first model can process information about a task and output a input for a second model to process in order to perform a step of the task. The second model can be the same as or different from the first model. Workbench 15 can implement prompt generation pipelines in development model 16.

[0156] Prompt libraries 17-4 can include pipelines for context injection. For instance, a performance of development model 16 on a particular task can improve if provided with additional context for performing the task. Prompt libraries 17-4 can include software components configured to identify desired context, retrieve the context from an external source (e.g., a database, a sensor, etc.), and add the context to the input prompt. Workbench 15 can implement context injection pipelines in development model 16.

[0157] Although various training examples described herein with respect to model development platform 12 refer to “pre-training” and “fine-tuning,” it is to be understood that model alignment toolkit 17 can generally support a wide variety of training techniques adapted for training a wide variety of machine-learned models. Example training techniques can correspond to the example training method 800 described above.

[0158] Model development platform 12 can include a model plugin toolkit 18. Model plugin toolkit 18 can include a variety of tools configured for augmenting the functionality of a machine-learned model by integrating the machine-learned model with other systems, devices, and software components. For instance, a machine-learned model can use tools to increase performance quality where appropriate. For instance, deterministic tasks can be offloaded to dedicated tools in lieu of probabilistically performing the task with an increased risk of error. For instance, instead of autoregressively predicting the solution to a system of equations, a machine-learned model can recognize a tool to call for obtaining the solution and pass the system of equations to the appropriate tool. The tool can be a traditional system of equations solver that can operate deterministically to resolve the system of equations. The output of the tool can be returned in response to the original query. In this manner, tool use can allow some example models to focus on the strengths of machine-learned models — e.g., understanding an intent in an unstructured request for a task — while augmenting the performance of the model by offloading certain tasks to a more focused tool for rote application of deterministic algorithms to a well-defined problem.

[0159] Model plugin toolkit 18 can include validation tools 18-1. Validation tools 18- 1 can include tools that can parse and confirm output(s) of a machine-learned model. Validation tools 18-1 can include engineered heuristics that establish certain thresholds applied to model outputs. For example, validation tools 18-1 can ground the outputs of machine-learned models to structured data sources (e.g., to mitigate “hallucinations”).

[0160] Model plugin toolkit 18 can include tooling packages 18-2 for implementing one or more tools that can include scripts or other executable code that can be executed alongside development model 16. Tooling packages 18-2 can include one or more inputs configured to cause machine-learned model(s) to implement the tools (e.g., few-shot prompts that induce a model to output tool calls in the proper syntax, etc.). Tooling packages 18-2 can include, for instance, fine-tuning training data for training a model to use a tool.

[0161] Model plugin toolkit 18 can include interfaces for calling external application programming interfaces (APIs) 18-3. For instance, in addition to or in lieu of implementing tool calls or tool code directly with development model 16, development model 16 can bealigned to output instructions that initiate API calls to send or obtain data via external systems.

[0162] Model plugin toolkit 18 can integrate with prompt libraries 17-4 to build a catalog of available tools for use with development model 16. For instance, a model can receive, in an input, a catalog of available tools, and the model can generate an output that selects a tool from the available tools and initiates a tool call for using the tool.

[0163] Model development platform 12 can include a computational optimization toolkit 19 for optimizing a computational performance of development model 16. For instance, tools for model compression 19-1 can allow development model 16 to be reduced in size while maintaining a desired level of performance. For instance, model compression 19-1 can include quantization workflows, weight pruning and sparsification techniques, etc. Tools for hardware acceleration 19-2 can facilitate the configuration of the model storage and execution formats to operate optimally on different hardware resources. For instance, hardware acceleration 19-2 can include tools for optimally sharding models for distributed processing over multiple processing units for increased bandwidth, lower unified memory requirements, etc. Tools for distillation 19-3 can provide for the training of lighter-weight models based on the knowledge encoded in development model 16. For instance, development model 16 can be a highly performant, large machine-learned model optimized using model development platform 12. To obtain a lightweight model for running in resource-constrained environments, a smaller model can be a “student model” that learns to imitate development model 16 as a “teacher model.” In this manner, for instance, the investment in learning the parameters and configurations of development model 16 can be efficiently transferred to a smaller model for more efficient inference.

[0164] Workbench 15 can implement one, multiple, or none of the toolkits implemented in model development platform 12. Workbench 15 can output an output model 20 based on development model 16. Output model 20 can be a deployment version of development model 16. Output model 20 can be a development or training checkpoint of development model 16. Output model 20 can be a distilled, compressed, or otherwise optimized version of development model 16.

[0165] Figure 13 is a block diagram of an example training flow for training a machine-learned development model 16. One or more portion(s) of the example training flow can be implemented by a computing system that includes one or more computing devices such as, for example, computing systems described with reference to the other figures. Each respective portion of the example training flow can be performed by any (or anycombination) of one or more computing devices. Moreover, one or more portion(s) of the example training flow can be implemented on the hardware components of the device(s) described herein, for example, to train one or more systems or models. FIG. 13 depicts elements performed in a particular order for purposes of illustration and discussion. Those of ordinary skill in the art, using the disclosures provided herein, will understand that the elements of any of the methods discussed herein can be adapted, rearranged, expanded, omitted, combined, or modified in various ways without deviating from the scope of the present disclosure. FIG. 13 is described with reference to elements / terms described with respect to other systems and figures for exemplary illustrated purposes and is not meant to be limiting. One or more portions of the example training flow can be performed additionally, or alternatively, by other systems.

[0166] Initially, development model 16 can persist in an initial state as an initialized model 21. Development model 16 can be initialized with weight values. Initial weight values can be random or based on an initialization schema. Initial weight values can be based on prior pre-training for the same or for a different model.

[0167] Initialized model 21 can undergo pre-training in a pre-training stage 22. Pretraining stage 22 can be implemented using one or more pre-training pipelines 17-2 over data from dataset(s) 17-1. Pre-training can be omitted, for example, if initialized model 21 is already pre-trained (e.g., development model 16 contains, is, or is based on a pre-trained foundational model or an expert model).

[0168] Pre-trained model 23 can then be a new version of development model 16, which can persist as development model 16 or as a new development model. Pre-trained model 23 can be the initial state if development model 16 was already pre-trained. Pre-trained model 23 can undergo fine-tuning in a fine-tuning stage 24. Fine-tuning stage 24 can be implemented using one or more fine-tuning pipelines 17-3 over data from dataset(s) 17-1. Fine-tuning can be omitted, for example, if a pre-trained model has satisfactory performance, if the model was already fine-tuned, or if other tuning approaches are preferred.

[0169] Fine-tuned model 25 can then be a new version of development model 16, which can persist as development model 16 or as a new development model. Fine-tuned model 25 can be the initial state if development model 16 was already fine-tuned. Fine-tuned model 25 can undergo refinement with user feedback 26. For instance, refinement with user feedback 26 can include reinforcement learning, optionally based on human feedback from human users of fine-tuned model 25. As reinforcement learning can be a form of fine-tuning, it is to be understood that fine-tuning stage 24 can subsume the stage for refining with userfeedback 26. Refinement with user feedback 26 can produce a refined model 27. Refined model 27 can be output to downstream system(s) 28 for deployment or further development.

[0170] In some implementations, computational optimization operations can be applied before, during, or after each stage. For instance, initialized model 21 can undergo computational optimization 29-1 (e.g., using computational optimization toolkit 19) before pre-training stage 22. Pre-trained model 23 can undergo computational optimization 29-2 (e.g., using computational optimization toolkit 19) before fine-tuning stage 24. Fine-tuned model 25 can undergo computational optimization 29-3 (e.g., using computational optimization toolkit 19) before refinement with user feedback 26. Refined model 27 can undergo computational optimization 29-4 (e.g., using computational optimization toolkit 19) before output to downstream system(s) 28. Computational optimization(s) 29-1, . . . , 29-4 can all be the same, all be different, or include at least some different optimization techniques.Example Machine-Learned Model Inference System

[0171] Figure 14 is a block diagram of an inference system for operating one or more machine-learned model(s) 1 to perform inference (e.g., for training, for deployment, etc.). A model host 31 can receive machine-learned model(s) 1. Model host 31 can host one or more model instance(s) 31-1, which can be one or multiple instances of one or multiple models. Model host 31 can host model instance(s) 31-1 using available compute resources 31-2 associated with model host 31.

[0172] Model host 31 can perform inference on behalf of one or more client(s) 32. Client(s) 32 can transmit an input request 33 to model host 31. Using input request 33, model host 31 can obtain input(s) 2 for input to machine-learned model(s) 1. Machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3. Using output(s) 3, model host 31 can return an output payload 34 for responding to input request 33 from client(s) 32. Output payload 34 can include or be based on output(s) 3.

[0173] Model host 31 can leverage various other resources and tools to augment the inference task. For instance, model host 31 can communicate with tool interfaces 35 to facilitate tool use by model instance(s) 31-1. Tool interfaces 35 can include local or remote APIs. Tool interfaces 35 can include integrated scripts or other software functionality. Model host 31 can engage online learning interface(s) 36 to facilitate ongoing improvements to machine-learned model(s) 1. For instance, online learning interface(s) 36 can be used within reinforcement learning loops to retrieve user feedback on inferences served by model host 31.Model host 31 can access runtime data source(s) 37 for augmenting input(s) 2 with additional contextual information. For instance, runtime data source(s) 37 can include a knowledge graph 37-1 that facilitates structured information retrieval for information associated with input request(s) 33 (e.g., a search engine service). Runtime data source(s) 37 can include public or private, external or local database(s) 37-2 that can store information associated with input request(s) 33 for augmenting input(s) 2. Runtime data source(s) 37 can include account data 37-3 which can be retrieved in association with a user account corresponding to a client 32 for customizing the behavior of model host 31 accordingly.

[0174] Model host 31 can be implemented by one or multiple computing devices or systems. Client(s) 2 can be implemented by one or multiple computing devices or systems, which can include computing devices or systems shared with model host 31.

[0175] For example, model host 31 can operate on a server system that provides a machine-learning service to client device(s) that operate client(s) 32 (e.g., over a local or wide-area network). Client device(s) can be end-user devices used by individuals. Client device(s) can be server systems that operate client(s) 32 to provide various functionality as a service to downstream end-user devices.

[0176] In some implementations, model host 31 can operate on a same device or system as client(s) 32. Model host 31 can be a machine-learning service that runs on-device to provide machine-learning functionality to one or multiple applications operating on a client device, which can include an application implementing client(s) 32. Model host 31 can be a part of a same application as client(s) 32. For instance, model host 31 can be a subroutine or method implemented by one part of an application, and client(s) 32 can be another subroutine or method that engages model host 31 to perform inference functions within the application. It is to be understood that model host 31 and client(s) 32 can have various different configurations.

[0177] Model instance(s) 31-1 can include one or more machine-learned models that are available for performing inference. Model instance(s) 31-1 can include weights or other model components that are stored in persistent storage, temporarily cached, or loaded into high-speed memory. Model instance(s) 31-1 can include multiple instance(s) of the same model (e.g., for parallel execution of more requests on the same model). Model instance(s) 31-1 can include instance(s) of different model(s). Model instance(s) 31-1 can include cached intermediate states of active or inactive model(s) used to accelerate inference of those models. For instance, an inference session with a particular model may generate significant amounts of computational results that can be re-used for future inference runs (e.g., using aKV cache for transformer-based models). These computational results can be saved in association with that inference session so that session can be executed more efficiently when resumed.

[0178] Compute resource(s) 31-2 can include one or more processors (central processing units, graphical processing units, tensor processing units, machine-learning accelerators, etc.) connected to one or more memory devices. Compute resource(s) 31-2 can include a dynamic pool of available resources shared with other processes. Compute resource(s) 31-2 can include memory devices large enough to fit an entire model instance in a single memory instance. Compute resource(s) 31-2 can also shard model instance(s) across multiple memory devices (e.g., using data parallelization or tensor parallelization, etc.). This can be done to increase parallelization or to execute a large model using multiple memory devices which individually might not be able to fit the entire model into memory.

[0179] Input request 33 can include data for input(s) 2. Model host 31 can process input request 33 to obtain input(s) 2. Input(s) 2 can be obtained directly from input request 33 or can be retrieved using input request 33. Input request 33 can be submitted to model host 31 via an API.

[0180] Model host 31 can perform inference over batches of input requests 33 in parallel. For instance, a model instance 31-1 can be configured with an input structure that has a batch dimension. Separate input(s) 2 can be distributed across the batch dimension (e.g., rows of an array). The separate input(s) 2 can include completely different contexts. The separate input(s) 2 can be multiple inference steps of the same task. The separate input(s) 2 can be staggered in an input structure, such that any given inference cycle can be operating on different portions of the respective input(s) 2. In this manner, for instance, model host 31 can perform inference on the batch in parallel, such that output(s) 3 can also contain the batch dimension and return the inference results for the batched input(s) 2 in parallel. In this manner, for instance, batches of input request(s) 33 can be processed in parallel for higher throughput of output payload(s) 34.

[0181] Output payload 34 can include or be based on output(s) 3 from machine- learned model(s) 1. Model host 31 can process output(s) 3 to obtain output payload 34. This can include chaining multiple rounds of inference (e.g., iteratively, recursively, across the same model(s) or different model(s)) to arrive at a final output for a task to be returned in output payload 34. Output payload 34 can be transmitted to client(s) 32 via an API.

[0182] Online learning interface(s) 36 can facilitate reinforcement learning of machine-learned model(s) 1. Online learning interface(s) 36 can facilitate reinforcementlearning with human feedback (RLHF). Online learning interface(s) 36 can facilitate federated learning of machine-learned model(s) 1.

[0183] Model host 31 can execute machine-learned model(s) 1 to perform inference for various tasks using various types of data. For example, various different input(s) 2 and output(s) 3 can be used for various different tasks. In some implementations, input(s) 2 can be or otherwise represent image data. Machine-learned model(s) 1 can process the image data to generate an output. As an example, machine-learned model(s) 1 can process the image data to generate an image recognition output (e.g., a recognition of the image data, a latent embedding of the image data, an encoded representation of the image data, a hash of the image data, etc.). As another example, machine-learned model(s) 1 can process the image data to generate an image segmentation output. As another example, machine-learned model(s) 1 can process the image data to generate an image classification output. As another example, machine-learned model(s) 1 can process the image data to generate an image data modification output (e.g., an alteration of the image data, etc.). As another example, machine- learned model(s) 1 can process the image data to generate an encoded image data output (e.g., an encoded and / or compressed representation of the image data, etc.). As another example, machine-learned model(s) 1 can process the image data to generate an upscaled image data output. As another example, machine-learned model(s) 1 can process the image data to generate a prediction output.

[0184] In some implementations, the task is a computer vision task. In some cases, input(s) 2 includes pixel data for one or more images and the task is an image processing task. For example, the image processing task can be image classification, where the output is a set of scores, each score corresponding to a different object class and representing the likelihood that the one or more images depict an object belonging to the object class. The image processing task may be object detection, where the image processing output identifies one or more regions in the one or more images and, for each region, a likelihood that region depicts an object of interest. As another example, the image processing task can be image segmentation, where the image processing output defines, for each pixel in the one or more images, a respective likelihood for each category in a predetermined set of categories. For example, the set of categories can be foreground and background. As another example, the set of categories can be object classes. As another example, the image processing task can be depth estimation, where the image processing output defines, for each pixel in the one or more images, a respective depth value. As another example, the image processing task can be motion estimation, where the network input includes multiple images, and the imageprocessing output defines, for each pixel of one of the input images, a motion of the scene depicted at the pixel between the images in the network input.

[0185] In some implementations, input(s) 2 can be or otherwise represent natural language data. Machine-learned model(s) 1 can process the natural language data to generate an output. As an example, machine-learned model(s) 1 can process the natural language data to generate a language encoding output. As another example, machine-learned model(s) 1 can process the natural language data to generate a latent text embedding output. As another example, machine-learned model(s) 1 can process the natural language data to generate a translation output. As another example, machine-learned model(s) 1 can process the natural language data to generate a classification output. As another example, machine-learned model(s) 1 can process the natural language data to generate a textual segmentation output. As another example, machine-learned model(s) 1 can process the natural language data to generate a semantic intent output. As another example, machine-learned model(s) 1 can process the natural language data to generate an upscaled text or natural language output (e.g., text or natural language data that is higher quality than the input text or natural language, etc.). As another example, machine-learned model(s) 1 can process the natural language data to generate a prediction output (e.g., one or more predicted next portions of natural language content).

[0186] In some implementations, input(s) 2 can be or otherwise represent speech data (e.g., data describing spoken natural language, such as audio data, textual data, etc.). Machine-learned model(s) 1 can process the speech data to generate an output. As an example, machine-learned model(s) 1 can process the speech data to generate a speech recognition output. As another example, machine-learned model(s) 1 can process the speech data to generate a speech translation output. As another example, machine-learned model(s) 1 can process the speech data to generate a latent embedding output. As another example, machine-learned model(s) 1 can process the speech data to generate an encoded speech output (e.g., an encoded and / or compressed representation of the speech data, etc.). As another example, machine-learned model(s) 1 can process the speech data to generate an upscaled speech output (e.g., speech data that is higher quality than the input speech data, etc.). As another example, machine-learned model(s) 1 can process the speech data to generate a textual representation output (e.g., a textual representation of the input speech data, etc.). As another example, machine-learned model(s) 1 can process the speech data to generate a prediction output.

[0187] In some implementations, input(s) 2 can be or otherwise represent latent encoding data (e.g., a latent space representation of an input, etc.). Machine-learned model(s) 1 can process the latent encoding data to generate an output. As an example, machine- learned model(s) 1 can process the latent encoding data to generate a recognition output. As another example, machine-learned model(s) 1 can process the latent encoding data to generate a reconstruction output. As another example, machine-learned model(s) 1 can process the latent encoding data to generate a search output. As another example, machine- learned model(s) 1 can process the latent encoding data to generate a reclustering output. As another example, machine-learned model(s) 1 can process the latent encoding data to generate a prediction output.

[0188] In some implementations, input(s) 2 can be or otherwise represent statistical data. Statistical data can be, represent, or otherwise include data computed and / or calculated from some other data source. Machine-learned model(s) 1 can process the statistical data to generate an output. As an example, machine-learned model(s) 1 can process the statistical data to generate a recognition output. As another example, machine-learned model(s) 1 can process the statistical data to generate a prediction output. As another example, machine- learned model(s) 1 can process the statistical data to generate a classification output. As another example, machine-learned model(s) 1 can process the statistical data to generate a segmentation output. As another example, machine-learned model(s) 1 can process the statistical data to generate a visualization output. As another example, machine-learned model(s) 1 can process the statistical data to generate a diagnostic output.

[0189] In some implementations, input(s) 2 can be or otherwise represent sensor data. Machine-learned model(s) 1 can process the sensor data to generate an output. As an example, machine-learned model(s) 1 can process the sensor data to generate a recognition output. As another example, machine-learned model(s) 1 can process the sensor data to generate a prediction output. As another example, machine-learned model(s) 1 can process the sensor data to generate a classification output. As another example, machine-learned model(s) 1 can process the sensor data to generate a segmentation output. As another example, machine-learned model(s) 1 can process the sensor data to generate a visualization output. As another example, machine-learned model(s) 1 can process the sensor data to generate a diagnostic output. As another example, machine-learned model(s) 1 can process the sensor data to generate a detection output.

[0190] In some implementations, machine-learned model(s) 1 can be configured to perform a task that includes encoding input data for reliable and / or efficient transmission orstorage (and / or corresponding decoding). For example, the task may be an audio compression task. The input may include audio data and the output may comprise compressed audio data. In another example, the input includes visual data (e.g. one or more images or videos), the output comprises compressed visual data, and the task is a visual data compression task. In another example, the task may comprise generating an embedding for input data (e.g. input audio or visual data). In some cases, the input includes audio data representing a spoken utterance and the task is a speech recognition task. The output may comprise a text output which is mapped to the spoken utterance. In some cases, the task comprises encrypting or decrypting input data. In some cases, the task comprises a microprocessor performance task, such as branch prediction or memory address translation.

[0191] In some implementations, the task is a generative task, and machine-learned model(s) 1 can be configured to output content generated in view of input(s) 2. For instance, input(s) 2 can be or otherwise represent data of one or more modalities that encodes context for generating additional content.

[0192] In some implementations, the task can be a text completion task. Machine- learned model(s) 1 can be configured to process input(s) 2 that represent textual data and to generate output(s) 3 that represent additional textual data that completes a textual sequence that includes input(s) 2. For instance, machine-learned model(s) 1 can be configured to generate output(s) 3 to complete a sentence, paragraph, or portion of text that follows from a portion of text represented by input(s) 2.

[0193] In some implementations, the task can be an instruction following task. Machine-learned model(s) 1 can be configured to process input(s) 2 that represent instructions to perform a function and to generate output(s) 3 that advance a goal of satisfying the instruction function (e.g., at least a step of a multi-step procedure to perform the function). Output(s) 3 can represent data of the same or of a different modality as input(s) 2. For instance, input(s) 2 can represent textual data (e.g., natural language instructions for a task to be performed) and machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data responsive to the instructions (e.g., natural language responses, programming language responses, machine language responses, etc.). Input(s) 2 can represent image data (e.g., image-based instructions for a task to be performed, optionally accompanied by textual instructions) and machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data responsive to the instructions (e.g., natural language responses, programming language responses, machine language responses, etc.). One or more output(s) 3 can be iteratively or recursively generated to sequentially processand accomplish steps toward accomplishing the requested functionality. For instance, an initial output can be executed by an external system or be processed by machine-learned model(s) 1 to complete an initial step of performing a function. Multiple steps can be performed, with a final output being obtained that is responsive to the initial instructions.

[0194] In some implementations, the task can be a question answering task. Machine- learned model(s) 1 can be configured to process input(s) 2 that represent a question to answer and to generate output(s) 3 that advance a goal of returning an answer to the question (e.g., at least a step of a multi-step procedure to perform the function). Output(s) 3 can represent data of the same or of a different modality as input(s) 2. For instance, input(s) 2 can represent textual data (e.g., natural language instructions for a task to be performed) and machine- learned model(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data responsive to the question (e.g., natural language responses, programming language responses, machine language responses, etc.). Input(s) 2 can represent image data (e.g., image-based instructions for a task to be performed, optionally accompanied by textual instructions) and machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data responsive to the question (e.g., natural language responses, programming language responses, machine language responses, etc.). One or more output(s) 3 can be iteratively or recursively generated to sequentially process and accomplish steps toward answering the question. For instance, an initial output can be executed by an external system or be processed by machine-learned model(s) 1 to complete an initial step of obtaining an answer to the question (e.g., querying a database, performing a computation, executing a script, etc.). Multiple steps can be performed, with a final output being obtained that is responsive to the question.

[0195] In some implementations, the task can be an image generation task. Machine- learned model(s) 1 can be configured to process input(s) 2 that represent context regarding a desired portion of image content. The context can include text data, image data, audio data, etc. Machine-learned model(s) 1 can be configured to generate output(s) 3 that represent image data that depicts imagery related to the context. For instance, machine-learned model(s) 1 can be configured to generate pixel data of an image. Values for channel(s) associated with the pixels in the pixel data can be selected based on the context (e.g., based on a probability determined based on the context).

[0196] In some implementations, the task can be an audio generation task. Machine- learned model(s) 1 can be configured to process input(s) 2 that represent context regarding a desired portion of audio content. The context can include text data, image data, audio data,etc. Machine-learned model(s) 1 can be configured to generate output(s) 3 that represent audio data related to the context. For instance, machine-learned model(s) 1 can be configured to generate waveform data in the form of an image (e.g., a spectrogram). Values for channel(s) associated with pixels of the image can be selected based on the context. Machine- learned model(s) 1 can be configured to generate waveform data in the form of a sequence of discrete samples of a continuous waveform. Values of the sequence can be selected based on the context (e.g., based on a probability determined based on the context).

[0197] In some implementations, the task can be a data generation task. Machine- learned model(s) 1 can be configured to process input(s) 2 that represent context regarding a desired portion of data (e.g., data from various data domains, such as sensor data, image data, multimodal data, statistical data, etc.). The desired data can be, for instance, synthetic data for training other machine-learned models. The context can include arbitrary data type(s). Machine-learned model(s) 1 can be configured to generate output(s) 3 that represent data that aligns with the desired data. For instance, machine-learned model(s) 1 can be configured to generate data values for populating a dataset. Values for the data object(s) can be selected based on the context (e.g., based on a probability determined based on the context).Example Computing Systems and Devices

[0198] Figure 15 is a block diagram of an example networked computing system that can perform aspects of example implementations of the present disclosure. The system can include a number of computing devices and systems that are communicatively coupled over a network 49. An example computing device 50 is described to provide an example of a computing device that can perform any aspect of the present disclosure (e.g., implementing model host 31, client(s) 32, or both). An example server computing system 60 is described as an example of a server computing system that can perform any aspect of the present disclosure (e.g., implementing model host 31, client(s) 32, or both). Computing device 50 and server computing system(s) 60 can cooperatively interact (e.g., over network 49) to perform any aspect of the present disclosure (e.g., implementing model host 31, client(s) 32, or both). Model development platform system 70 is an example system that can host or serve model development platform(s) 12 for development of machine-learned models. Third-party system(s) 80 are example system(s) with which any of computing device 50, server computing system(s) 60, or model development platform system(s) 70 can interact in the performance of various aspects of the present disclosure (e.g., engaging third-party tools, accessing third-party databases or other resources, etc.).

[0199] Network 49 can be any type of communications network, such as a local area network (e.g., intranet), wide area network (e.g., Internet), or some combination thereof and can include any number of wired or wireless links. In general, communication over network 49 can be carried via any type of wired or wireless connection, using a wide variety of communication protocols (e.g., TCP / IP, HTTP, SMTP, FTP), encodings or formats (e.g., HTML, XML), or protection schemes (e.g., VPN, secure HTTP, SSL). Network 49 can also be implemented via a system bus. For instance, one or more devices or systems of Figure 15 can be co-located with, contained by, or otherwise integrated into one or more other devices or systems.

[0200] Computing device 50 can be any type of computing device, such as, for example, a personal computing device (e.g., laptop or desktop), a mobile computing device (e.g., smartphone or tablet), a gaming console or controller, a wearable computing device, an embedded computing device, a server computing device, a virtual machine operating on a host device, or any other type of computing device. Computing device 50 can be a client computing device. Computing device 50 can be an end-user computing device. Computing device 50 can be a computing device of a service provided that provides a service to an end user (who may use another computing device to interact with computing device 50).

[0201] Computing device 50 can include one or more processors 51 and a memory 52. Processor(s) 51 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. Memory 52 can include one or more non-transitory computer-readable storage media, such as HBM, RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. Memory 52 can store data 53 and instructions 54 which can be executed by processor(s) 51 to cause computing device 50 to perform operations. The operations can implement any one or multiple features described herein. The operations can implement example methods and techniques described herein.

[0202] Computing device 50 can also include one or more input components that receive user input. For example, a user input component can be a touch-sensitive component (e.g., a touch-sensitive display screen or a touch pad) that is sensitive to the touch of a user input object (e.g., a finger or a stylus). The touch-sensitive component can serve to implement a virtual keyboard. Other example user input components include a microphone, camera, LIDAR, a physical keyboard or other buttons, or other means by which a user can provide user input.

[0203] Computing device 50 can store or include one or more machine-learned models 55. Machine-learned models 55 can include one or more machine-learned model(s) 1, such as a sequence processing model 4. Machine-learned models 55 can include one or multiple model instance(s) 31-1. Machine-learned model(s) 55 can be received from server computing system(s) 60, model development platform system 70, third party system(s) 80 (e.g., an application distribution platform), or developed locally on computing device 50. Machine-learned model(s) 55 can be loaded into memory 52 and used or otherwise implemented by processor(s) 51. Computing device 50 can implement multiple parallel instances of machine-learned model(s) 55.

[0204] Server computing system(s) 60 can include one or more processors 61 and a memory 62. Processor(s) 61 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. Memory 62 can include one or more non-transitory computer-readable storage media, such as HBM, RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. Memory 62 can store data 63 and instructions 64 which can be executed by processor(s) 61 to cause server computing system(s) 60 to perform operations. The operations can implement any one or multiple features described herein. The operations can implement example methods and techniques described herein.

[0205] In some implementations, server computing system 60 includes or is otherwise implemented by one or multiple server computing devices. In instances in which server computing system 60 includes multiple server computing devices, such server computing devices can operate according to sequential computing architectures, parallel computing architectures, or some combination thereof.

[0206] Server computing system 60 can store or otherwise include one or more machine-learned models 65. Machine-learned model(s) 65 can be the same as or different from machine-learned model(s) 55. Machine-learned models 65 can include one or more machine-learned model(s) 1, such as a sequence processing model 4. Machine-learned models 65 can include one or multiple model instance(s) 31-1. Machine-learned model(s) 65 can be received from computing device 50, model development platform system 70, third party system(s) 80, or developed locally on server computing system(s) 60. Machine-learned model(s) 65 can be loaded into memory 62 and used or otherwise implemented by processor(s) 61. Server computing system(s) 60 can implement multiple parallel instances of machine-learned model (s) 65.

[0207] In an example configuration, machine-learned models 65 can be included in or otherwise stored and implemented by server computing system 60 to establish a client-server relationship with computing device 50 for serving model inferences. For instance, server computing system(s) 60 can implement model host 31 on behalf of client(s) 32 on computing device 50. For instance, machine-learned models 65 can be implemented by server computing system 60 as a portion of a web service (e.g., remote machine-learned model hosting service, such as an online interface for performing machine-learned model operations over a network on server computing system(s) 60). For instance, server computing system(s) 60 can communicate with computing device 50 over a local intranet or internet connection. For instance, computing device 50 can be a workstation or endpoint in communication with server computing system(s) 60, with implementation of machine-learned models 65 being managed by server computing system(s) 60 to remotely perform inference (e.g., for runtime or training operations), with output(s) returned (e.g., cast, streamed, etc.) to computing device 50. Machine-learned models 65 can work cooperatively or interoperatively with machine- learned models 55 on computing device 50 to perform various tasks.

[0208] Model development platform system(s) 70 can include one or more processors 71 and a memory 72. Processor(s) 71 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. Memory 72 can include one or more non-transitory computer-readable storage media, such as HBM, RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. Memory 72 can store data 73 and instructions 74 which can be executed by processor(s) 71 to cause model development platform system(s) 70 to perform operations. The operations can implement any one or multiple features described herein. The operations can implement example methods and techniques described herein. Example operations include the functionality described herein with respect to model development platform 12. This and other functionality can be implemented by developer tool(s) 75.

[0209] Third-party system(s) 80 can include one or more processors 81 and a memory 82. Processor(s) 81 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. Memory 82 can include one or more non-transitory computer-readable storage media, such as HBM, RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. Memory 82 can store data 83 and instructions 84 which can be executed by processor(s) 81 tocause third-party system(s) 80 to perform operations. The operations can implement any one or multiple features described herein. The operations can implement example methods and techniques described herein. Example operations include the functionality described herein with respect to tools and other external resources called when training or performing inference with machine-learned model(s) 1, 4, 16, 20, 55, 65, etc. (e.g., third-party resource(s) 85).

[0210] Figure 15 illustrates one example arrangement of computing systems that can be used to implement the present disclosure. Other computing system configurations can be used as well. For example, in some implementations, one or both of computing system 50 or server computing system(s) 60 can implement all or a portion of the operations of model development platform system 70. For example, computing system 50 or server computing system(s) 60 can implement developer tool(s) 75 (or extensions thereof) to develop, update / train, or refine machine-learned models 1, 4, 16, 20, 55, 65, etc. using one or more techniques described herein with respect to model alignment toolkit 17. In this manner, for instance, computing system 50 or server computing system(s) 60 can develop, update / train, or refine machine-learned models based on local datasets (e.g., for model personalization / customization, as permitted by user data preference selections).

[0211] Figure 16 is a block diagram of an example computing device 98 that performs according to example embodiments of the present disclosure. Computing device 98 can be a user computing device or a server computing device (e.g., computing device 50, server computing system(s) 60, etc.). Computing device 98 can implement model host 31. For instance, computing device 98 can include a number of applications (e.g., applications 1 through N). Each application can contain its own machine learning library and machine- learned model(s). For example, each application can include a machine-learned model. Example applications include a text messaging application, an email application, a dictation application, a virtual keyboard application, a browser application, etc. As illustrated in Figure 16, each application can communicate with a number of other components of the computing device, such as, for example, one or more sensors, a context manager, a device state component, or additional components. In some implementations, each application can communicate with each device component using an API (e.g., a public API). In some implementations, the API used by each application is specific to that application.

[0212] Figure 17 is a block diagram of an example computing device 99 that performs according to example embodiments of the present disclosure. Computing device 99 can be the same as or different from computing device 98. Computing device 99 can be a usercomputing device or a server computing device (e.g., computing device 50, server computing system(s) 60, etc.). Computing device 98 can implement model host 31. For instance, computing device 99 can include a number of applications (e.g., applications 1 through N). Each application can be in communication with a central intelligence layer. Example applications include a text messaging application, an email application, a dictation application, a virtual keyboard application, a browser application, etc. In some implementations, each application can communicate with the central intelligence layer (and model(s) stored therein) using an API (e.g., a common API across all applications).

[0213] The central intelligence layer can include a number of machine-learned models. For example, as illustrated in Figure 17, a respective machine-learned model can be provided for each application and managed by the central intelligence layer. In other implementations, two or more applications can share a single machine-learned model. For example, in some implementations, the central intelligence layer can provide a single model for all of the applications. In some implementations, the central intelligence layer is included within or otherwise implemented by an operating system of computing device 99.

[0214] The central intelligence layer can communicate with a central device data layer. The central device data layer can be a centralized repository of data for computing device 99. As illustrated in Figure 17, the central device data layer can communicate with a number of other components of the computing device, such as, for example, one or more sensors, a context manager, a device state component, or additional components. In some implementations, the central device data layer can communicate with each device component using an API (e.g., a private API).Additional Disclosure

[0215] The technology discussed herein makes reference to servers, databases, software applications, and other computer-based systems, as well as actions taken and information sent to and from such systems. The inherent flexibility of computer-based systems allows for a great variety of possible configurations, combinations, and divisions of tasks and functionality between and among components. For instance, processes discussed herein can be implemented using a single device or component or multiple devices or components working in combination. Databases and applications can be implemented on a single system or distributed across multiple systems. Distributed components can operate sequentially or in parallel.

[0216] While the present subject matter has been described in detail with respect to various specific example embodiments thereof, each example is provided by way of explanation, not limitation of the disclosure. Those skilled in the art, upon attaining an understanding of the foregoing, can readily produce alterations to, variations of, and equivalents to such embodiments. Accordingly, the subject disclosure does not preclude inclusion of such modifications, variations or additions to the present subject matter as would be readily apparent to one of ordinary skill in the art. For instance, features illustrated or described as part of one embodiment can be used with another embodiment to yield a still further embodiment. Thus, it is intended that the present disclosure cover such alterations, variations, and equivalents.

[0217] Aspects of the disclosure have been described in terms of illustrative embodiments thereof. Any and all features in the following claims can be combined or rearranged in any way possible, including combinations of claims not explicitly enumerated in combination together, as the example claim dependencies listed herein should not be read as limiting the scope of possible combinations of features disclosed herein. Accordingly, the scope of the present disclosure is by way of example rather than by way of limitation, and the subject disclosure does not preclude inclusion of such modifications, variations or additions to the present subject matter as would be readily apparent to one of ordinary skill in the art. Moreover, terms are described herein using lists of example elements joined by conjunctions such as “and,” “or,” “but,” etc. It should be understood that such conjunctions are provided for explanatory purposes only. Clauses and other sequences of items joined by a particular conjunction such as “or,” for example, can refer to “and / or,” “at least one of’, “any combination of’ example elements listed therein, etc. Terms such as “based on” should be understood as “based at least in part on.”

[0218] The term “can” should be understood as referring to a possibility of a feature in various implementations and not as prescribing an ability that is necessarily present in every implementation. For example, the phrase “X can perform Y” should be understood as indicating that, in various implementations, X has the potential to be configured to perform Y, and not as indicating that in every instance X must always be able to perform Y. It should be understood that, in various implementations, X might be unable to perform Y and remain within the scope of the present disclosure.

[0219] The term “may” should be understood as referring to a possibility of a feature in various implementations and not as prescribing an ability that is necessarily present in every implementation. For example, the phrase “X may perform Y” should be understood asindicating that, in various implementations, X has the potential to be configured to perform Y, and not as indicating that in every instance X must always be able to perform Y. It should be understood that, in various implementations, X might be unable to perform Y and remain within the scope of the present disclosure.

[0220] As used herein, any claim to an apparatus-implemented or deviceimplemented (e.g., computer-implemented) method describes a method that can be “used” (e.g., performed, executed, infringed, etc.) by a legal entity or legal person (e.g., corporation, individual human being, etc.) by initiating the method in any manner. For example, initiating can include causing, in any manner, one or more devices to perform the activities of the claimed method. The one or more devices can include, for example, one or more first devices owned by the legal entity or legal person initiating the method; one or more second devices not owned by the legal entity or legal person initiating the method; one or more third devices administered, maintained, or otherwise controlled by the legal entity or legal person initiating the method; and one or more fourth devices not administered, maintained or otherwise controlled by the legal entity or legal person initiating the method. Causing can include any form of causation, including but not limited to but-for causation, substantial factor causation such as independent-sufficient-cause causation, and the like. By way of non-limiting illustrative example, causing a computer to perform an action can include a variety of acts that can be performed by a legal entity, such as providing or causing to be provided a signal (e.g., network signal) or computer-readable instruction (e.g., API instruction, HTML instruction for execution by a browser application, software application, etc.) to cause the computer to perform the action; providing or causing to be provided an input interaction (e.g., user input interaction such as touch screen interaction, button press, keyboard data entry, etc.) to cause the computer to perform the action; installing or causing to be installed, or activating or causing to be activated, or providing or causing to be provided for installation or activation, one or more computer-readable instructions (e.g., in one or more non-transitory computer-readable storage media associated with the computer) that, when executed, cause the computer to perform the action; or any other causal action. Additionally, causing a device to perform an action can include causation (e.g., but-for causation) that may depend on one or more intervening causes. By way of non-limiting illustrative example, causing a computer to perform an action can include providing a first plurality of computer-readable instructions to cause the computer to: perform the claimed action using a second plurality of computer- readable instructions (e.g., web browser, etc.); perform the claimed action responsive to a predetermined user action (e.g., interacting with a “Click to continue” button; selecting a firstconfiguration option of a finite plurality of configuration options; opting into or not opting out of a predetermined protocol comprising the claimed action; etc.); or the like.

[0221] Additionally, as used herein, any claim to a device or system describes a device or system that can be “used” by a legal entity or legal person by putting the device or system into action in any manner. As a non-limiting illustrative example, as used herein, any claim to a device or system comprising non-transitory computer-readable media storing instructions that, when executed by one or more devices (e.g., processor devices, computing devices , etc.), cause the one or more devices to perform one or more claimed operations describes a system or device that can be “used” by causing, in any manner (e.g., any manner described above with respect to initiating performance of a method), one or more devices (e.g., devices owned or not owned by a legal person putting the devices into action, etc.) to perform the one or more claimed operations.

Claims

WHAT IS CLAIMED IS:

1. A method for training a machine-learned model using federated learning, comprising: providing, by a first computing device to a computing system comprising one or more second computing devices, data indicative of one or more input values for one or more layers of the machine-learned model; receiving, by the first computing device from the computing system, data indicative of a forward pass of the machine-learned model based on the one or more input values, wherein the data indicative of the forward pass comprises a reduced-parameter-count model description associated with the machine-learned model; and determining, by the first computing device based on the data indicative of the forward pass, one or more update values for the machine-learned model, wherein the one or more update values comprise at least one update value for at least one parameter of the reduced- parameter-count model description.

2. The method of claim 1, further comprising providing, by the first computing device, the one or more update values to the computing system.

3. The method of claim 1, further comprising updating, by the first computing device based on the one or more update values, one or more local parameters stored locally on the first computing device, wherein the local parameters are associated with the machine-learned model.

4. The method of claim 1, wherein: the first computing device is a first client device; the computing system comprises one or more server devices; and the forward pass was performed by the one or more server devices.

5. The method of claim 4, wherein the computing system further comprises one or more second client devices, and further comprising: performing, by the first computing device, at least one of: masking the one or more update values and subsequently providing a result of the masking to the one or more second client devices; andreceiving at least one masked update value from the one or more second client devices and providing one or more aggregated update values to the computing system based on the one or more update values and the at least one masked update value.

6. The method of claim 1, wherein the data indicative of the forward pass comprises data indicative of a strict subset of a plurality of parameters of a layer of the one or more layers.

7. The method of claim 6, wherein the strict subset consists of less than 50 percent of the plurality of parameters of the layer of the one or more layers.

8. The method of claim 6, wherein the strict subset consists of less than 10 percent of the plurality of parameters of the layer of the one or more layers.

9. The method of claim 1, wherein: the reduced-parameter-count model description comprises one or more first parameters of the one or more layers, the first parameters having a non-zero-valued contribution to one or more output values of a final layer of the one or more layers during the forward pass; and the reduced-parameter-count model description omits one or more second parameters of the one or more layers, the second parameters having a zero-valued contribution to the one or more output values during the forward pass.

10. The method of claim 9, wherein: determining the one or more update values comprises computing, based at least in part on the one or more first parameters, a gradient of an objective function associated with the machine-learned model; and the objective function is based at least in part on the one or more output values.

11. The method of claim 1, wherein the data indicative of the forward pass comprises a plurality of output activation values of a plurality of nodes of the one or more layers.

12. The method of claim 1, wherein data indicative of the one or more input values comprise data indicative of a user input received from a user of the first computing device,the user input comprising one or more of: text-based input, audio input, image input, and video input.

13. The method of claim 1, wherein the one or more update values is determined based on private data not available to the computing system.

14. The method of claim 1, wherein the machine-learned model comprises a sequence processing model.

15. A computing system for training a machine-learned model using federated learning, comprising: one or more computing devices comprising one or more processors and one or more non-transitory computer-readable media storing instructions that are executable by the one or more computing devices to perform operations, the operations comprising: receiving, from a first computing device, data indicative of one or more input values for one or more layers of the machine-learned model; performing a forward pass of the machine-learned model based on the data indicative of the one or more input values; and providing, to the first computing device, data indicative of the forward pass, wherein the data indicative of the forward pass comprises a reduced-parameter-count model description of the machine-learned model.

16. The computing system of claim 15, wherein the operations further comprise: receiving, from the first computing device, data indicative of one or more update values for the machine-learned model; and updating the machine-learned model based at least in part on the data indicative of the update values.

17. The computing system of claim 15, wherein the data indicative of the forward pass comprises data indicative of a strict subset of a plurality of parameters of a layer of the one or more layers.

18. The computing system of claim 15, wherein:the reduced-parameter-count model description comprises one or more first parameters of the one or more layers, the first parameters having a non-zero-valued contribution to one or more output values of a final layer of the one or more layers during the forward pass; and the reduced-parameter-count model description omits one or more second parameters of the one or more layers, the second parameters having a zero-valued contribution to the one or more output values during the forward pass.

19. The computing system of claim 15, wherein the data indicative of the forward pass comprises a plurality of output activation values of a plurality of nodes of the one or more layers.

20. One or more non-transitory computer-readable media storing instructions that are executable by a first computing system to perform operations, the operations comprising: providing, to a second computing system comprising one or more computing devices, data indicative of one or more input values for one or more layers of a machine-learned model; receiving, from the second computing system, data indicative of a forward pass of the machine-learned model based on the one or more input values, wherein the data indicative of the forward pass comprises a reduced-parameter-count model description of the machine- learned model; and determining, based on the data indicative of the forward pass, one or more update values for the machine-learned model, wherein the one or more update values comprise at least one update value for at least one parameter of the reduced-parameter-count model description.

Images