Patent Application
20230177382
Federated learning of machine learning (ML) model(s) is an increasingly popular ML technique for training ML model(s). In traditional federated learning, an on-device ML model is stored locally on a client device of a user, and a global ML model, that is a remote-based counterpart of the on-device ML model, is stored remotely at a remote system (e.g., a remote server or a cluster of remote servers). The client device, using the on-device ML model, can process user input detected at the client device to generate predicted output, and can generate a gradient based on the predicted output in a supervised or unsupervised manner. Further, the client device can transmit the gradient to the remote system. The remote system can utilize the gradient, and optionally additional gradients generated in a similar manner at the client device or additional client devices, to update weights of the global ML model. The remote system can transmit the global ML model, or updated weights of the global ML model, to the client device. The client device can then replace the on-device ML model with the updated global ML model, or replace the weights of the on-device ML model with the updated weights of the updated global ML model, thereby updating the on-device ML model.
However, federated learning of ML model(s) may cause a large quantity of resources to be unnecessarily consumed at the client device and/or the client device may have insufficient resources to store and train full-sized ML model(s) locally at the client device. For example, assume an on-device automatic speech recognition (ASR) model is utilized to generate a transcription that is predicted to correspond to a spoken utterance captured in a stream of audio data via microphone(s) of the client device. In this example, the predicted output can correspond to the transcription, or a portion thereof, and the gradient can be generated based on the transcription. In generating the gradient, on-device memory may be consumed by storing activations at each layer of the on-device ASR model, and on-device computational resources may be consumed by computing an error function at each layer of the on-device ASR model and causing the error function to be backpropagated across each layer of the on-device ASR model. As a result, techniques that do not unnecessarily consume on-device memory and on-device computational resources are needed to improve efficiency of federated learning.
Implementations disclosed herein are directed to efficient federated learning of machine learning (ML) model(s) at a remote system (e.g., a remote server or cluster of remote servers) based on update(s) generated at client device(s). Processor(s) of a given client device may receive client data, process the client data using a given on-device ML model to generate predicted output, generate a gradient based on the predicted outputs using supervised or unsupervised learning techniques, generate a given update for a given disparate portion of the given on-device ML model based on the gradient, and transmit the given update to the remote system. Further, processor(s) of the remote system may receive the given update from the given client device, and cause a given global ML model (e.g., that is a remote-based counterpart of the given on-device ML model) to be updated based on the given update, thereby updating a given disparate portion of the given global ML model that corresponds to the given disparate portion of the given on-device ML model and without causing any other portions of the given ML model to be updated based on the given update. Notably, this process may be repeated at the given client device (or in parallel at one or more additional client devices) to generate additional given updates for other disparate portions of the given global ML model. As a result, computational resources consumed at the client devices that participate in the federated learning of the ML model(s) may be reduced since the client devices do not generate full updates for the full-size ML model(s), and network resources consumed in transmitting the updates to the remote system may be reduced since less than full updates for the full-size ML model(s) are transmitted to the remote system.
For example, assume audio data capturing a spoken utterance of a human user of “Hey Assistant, turn on the kitchen lights” is generated via microphone(s) of a given client device of the human user. In this example, the audio data capturing the spoken utterance may correspond to the client data. Further, the audio data may be processed using an on-device automatic speech recognition (ASR) model stored in on-device memory of the given client device to generate ASR data, such as one or more speech hypotheses that are predicted to correspond to the spoken utterance captured in the audio data. In this example the one or more speech hypotheses may correspond to the predicted output. Moreover, a gradient may be generated based on processing the audio data and/or one or more of the speech hypotheses using one or more semi-supervised or self-supervised learning techniques described herein. Notably, the on-device ASR model may include a plurality of on-device ASR layers, such as one or more first on-device ASR layers, one or more second on-device ASR layers, one or more third on-device ASR layers, and so on. However, the gradient may be generated for only a disparate portion of the on-device ASR model, such as the one or more first on-device ASR layers, but not for any of the other on-device ASR layers. By generating the gradient for only the one or more first on-device ASR layers, consumption of on-device memory and computational resources at the given client device may be reduced since the given client device does not have to store activation functions for each and every ASR layer of the on-device ASR model to generate the gradient and since the given client device does not have to compare the activation functions to a loss function for each and every ASR layer of the on-device ASR model as in traditional federated learning.
Continuing with the above example, assume the given client device generates a gradient for the one or more first on-device ASR layers. The given client device may further generate a first update for the one or more first on-device ASR layers based on the gradient. In this example, the first update may include, for instance, an indication of the one or more first on-device ASR layers of the on-device ASR model associated with the first update, an indication of one or more first counterpart global ASR layers of a counterpart global ASR model to be updated based on the first update, an indication of the on-device ASR model utilized to generate the first update for the one or more first on-device ASR layers of the on-device ASR model, the gradient generated based on the predicted output, one or more updated first on-device ASR layers of the on-device ASR model that are updated locally at the client device based on the gradient (e.g., in some implementations where the on-device ASR is, in fact, updated locally at the given client device based on the gradient), and/or one or more updated on-device weights for the one or more updated first on-device ASR layers of the on-device ASR model (e.g., in some implementations where the on-device ASR is, in fact, updated locally at the given client device based on the gradient).
In this example, the given client device may transmit the first update to the remote system. In response to receiving the first update, the remote system may cause the one or more first global ASR layers of the global ASR model to be updated based on the first update, resulting in an updated global ASR model. In some instances, the remote system may optionally combine the first update received from the given client device with additional updates received from the given client device or one or more additional client devices. Further, the remote system may transmit the updated global ASR model (e.g., the updated global ASR model itself, one or more updated global ASR layers thereof, and/or one or more updated global weights thereof) back to the client device and/or one or more of the additional client devices. Each of the client devices may replace, in corresponding on-device memory, the on-device ASR model with the updated global ASR model, and continue generating additional updates for the updated global ASR model, and transmitting the additional updates to the remote system to continue updating the updated global ASR model.
Although the above example is described with respect to the client data being audio data that captures a spoken utterance, it should be understood that is for the sake of example and is not meant to be limiting. For instance, the client data can additionally, or alternatively, include vision data, textual data, touch data, and/or any other type of client data generated at the given client device. In these and other instances, the on-device ML models utilized to process the client data may be vision-based ML model(s), text-based ML model(s), and/or touch-based ML model(s). Further, although the above example is described with respect to ASR models utilized to process the audio data, it should be understood that is also for the sake of example and is not meant to be limiting. For instance, other audio-based ML model(s) may additionally or alternatively be utilized to process the client data and be subsequently updated based on the processing of the client data. In these and other instances, the audio-based ML model(s) may include, for example, hotword detection models trained to predict whether the audio data captures a particular word or phrase, a voice activity detection (VAD) model trained to predict whether voice activity of a human is captured in the stream audio data, and/or any other audio-based ML model(s).
In some implementations, and prior to generating the gradient, multiple on-device ML layers may be compressed such that the multiple on-device ML layers are implemented as a single on-device ML layer at the given client device. For example, the given client device may compress a first on-device ML layer and a second on-device ML layer into one or more first on-device ML layers based on a scheduling signal generated by an on-device scheduling engine of the given client device and/or received from a remote scheduling engine. The scheduling signal may include an indication of the on-device ML layers to be compressed prior to generating the gradient. In some versions of those examples, the first on-device ML layer and the second on-device ML layer may be compressed into the one or more first on-device ML layers prior to generating the predicted output and prior to generating the gradient, whereas in additional or alternative examples, the first on-device ML layer and the second on-device ML layer may be compressed into the one or more first on-device ML layers subsequent to generating the predicted output but prior to generating the gradient. Notably, neither the first on-device ML layer nor the second on-device ML layer are dropped out. Rather, compressing the first on-device ML layer and the second on-device ML layer into the one or more first ML layers imposes a constraint on the on-device ML model.
Further, an additional client device, and prior to generating an additional gradient (e.g., based on additional client data received at the additional client device), multiple on-device ML layers may also be compressed such that the multiple on-device ML layers are implemented as a single on-device ML layer at the additional client device. Continuing with the above example, at least the second on-device ML layer and a third on-device ML layer of the on-device ML model into one or more second ML layers at the additional client device. The additional client device may compress the second on-device ML layer and the third on-device ML layer of the on-device ML model into the one or more second ML layers in the same or similar manner described above, but based on a scheduling signal generated by an on-device scheduling engine of the additional client device and/or received from a remote scheduling engine. Notably, in this example, the second on-device ML layer is compressed into both the one or more first on-device ML layers at the client device and the one or more second on-device ML layers at the additional client device. However, the one or more first on-device ML layers and the one or more second on-device ML layers each include at least one unique on-device ML layer—the first on-device ML layer for the one or more first on-device ML layers and the third on-device ML layer for the one or more second on-device ML layer. Nonetheless, it should be understood that this is for the sake of example and is not meant to be limiting. For instance, the same on-device ML layers can be compressed at disparate client devices (e.g., the first on-device ML layer and the second on-device ML layer being compressed into the one or more first ML layers at both the client device and the additional client device), disparate ML layers can be compressed at disparate client devices such that there is at least one unique ML layer compressed at the disparate devices, and/or disparate ML layers can be compressed at disparate client devices such that there is only unique ML layers compressed at the different devices. Although differences may exist between these disparate ML layers, they are adjacent in the on-device ML model and may include some overlap. Accordingly, by generating the first update and the second update as shared updates, the remote system can leverage this overlap to conserve computational and/or network resources as described herein, thereby resulting in improved efficiency in federated learning.
In some implementations, the remote scheduling engine may divide a superset of client devices into a plurality of subsets. In some versions of those implementations, each of the subsets of client devices may perform an iteration of a gradient snapshotting technique in updating the given global ML model. For example, the remote system may cause the given client device to generate the update for one or more Nth on-device ML layers of the given on-device ML model based on processing the client data received at the client device, where N is a positive integer. Further, the remote system may cause an additional client device to generate an additional update for one or more Nth−1 on-device ML layers of the given on-device ML model based on processing additional client data received at the additional client device. Put another way, the system may employ the gradient snapshotting technique to improve efficiency of federated learning of ML model(s). The gradient snapshotting technique may be implemented based on corresponding scheduling signals generated by a corresponding on-device scheduling engine of the client device and/or the additional client device and/or received from the remote scheduling engine. The corresponding scheduling signals may instruct the client device to generate the first update for the one or more Nth on-device ML layers of the on-device ML model, the additional client device to generate the second update for the one or more Nth−1 on-device ML layers of the on-device ML model, a further additional client device to generate a third update for one or more Nth−2 on-device ML layers of the on-device ML model, and so on. Accordingly, by utilizing the gradient snapshotting technique, the remote system can ensure that an update is generated for each of the Nth layers on-device ML model as described herein.
In some versions of those implementations, the remote scheduling engine may divide a superset of client devices into a plurality of subsets. In some versions of those implementations, each of the subsets of client devices may perform an iteration of the gradient snapshotting technique. For example, a first subset of client devices may include a first client device may generate an update for the one or more Nth on-device ML layers, a second client device may generate an update for the one or more Nth−1 on-device ML layers, a third client device may generate an update for the one or more Nth−2 on-device ML layers, and so on for each of the Nth on-device ML layers. Further, a second subset of client devices may generate an update for each of the Nth on-device ML layers in the same or similar manner. Accordingly, when the updates are sent to the remote system, it may be ensured that the remote system has updates for each of the disparate portions of ML model available for updating.
In some implementations, the unsupervised learning utilized to the gradient may correspond to a semi-supervised learning technique, such as a teacher-student approach. For example, the predicted output can be generated based on processing the client data using the on-device ML model as described in the above example. Further, benchmark output can be generated based on processing the client data using a benchmark ML model. In this example, the benchmark ML model may be of a same type as the on-device ML model and the global ML model, and the benchmark output may be utilized as a supervision signal for generating the gradient. For instance, a benchmark ML model may correspond to a benchmark ASR model that is utilized to generate one or more benchmark speech hypotheses based on processing audio data. Further, an on-device ASR model may generate one or more of the speech hypotheses based on processing the audio as described in the above example. In this instance, one or more of the benchmark speech hypotheses may be compared to one or more of the speech hypotheses to generate the gradient for the one or more first on-device ASR layers.
In some versions of those implementations, the predicted output may only be generated using the on-device ASR model and/or the gradient may only be utilized in generating the update in response to determining that one or more conditions are satisfied. The one or more conditions can include, for example, whether the predicted output satisfies a predicted output threshold, whether the benchmark output satisfies a benchmark output threshold, and/or other conditions. Put another way, the predicted output may only be generated using the on-device ASR model and/or the gradient may only be utilized in generating the update in response to determining that the benchmark output provides a sufficient supervision signal.
In additional or alternative implementations, the unsupervised learning utilized to generate the gradient can correspond to a self-supervised learning technique. In self-supervised learning, a target portion of client data may be identified. The target portion of the client data may be subsequent to a prepended portion of the client data and may be prior to an appended portion of the client data. Further, the target portion of the client data may be masked using various masking techniques. The target portion of the client data may be selected arbitrarily, or selected based on one or more criteria such as a particular segment between n and m seconds of audio data corresponds to the target portion, a portion of vision data corresponding to a center portion of an image corresponds to the target portion, a center portion of textual data corresponds to the target portion, and/or any other criteria for selecting the target portion of the client data. In some implementations when the client data corresponds to audio data, the target portion of the client data may correspond to a target audio waveform portion of the corresponding stream of audio data, the prepended portion of the client data may correspond to a prepended audio waveform portion that is received prior to the target audio waveform portion, and the appended portion of the client data may correspond to an appended audio waveform portion that is received subsequent to the target audio waveform portion. In additional or alternative implementations when the client data corresponds to audio data, the client data may correspond to a representation of the audio, such as a vector representation of the audio data and/or other representations of the audio data. In these implementations, the target portion of the client data may correspond to a target representation portion of the audio data, the prepended portion of the client data may correspond to a prepended representation portion that precedes the target representation portion, and the appended portion of the client data may correspond to an appended representation portion that follows subsequent to the target representation portion. In additional or alternative implementations, the client data may correspond to textual data or vision data. In these implementations, the target portion, the prepended portion, and the appended portion may be identified in the same or similar manner, and the target portion may be masked.
In implementations when the unsupervised learning utilized to the gradient corresponds to the self-supervised learning technique described above, the prepended portion of the client data and the appended portion of the client data may be processed using the on-device ML model to generate the predicted output that predict the target portion of the client data. For example, in implementations where the target portion of the client data corresponds to the target audio waveform portion of the corresponding stream of audio data, further assume the audio data captures a spoken utterance. In this example, the prepended audio waveform portion and the appended audio waveform portion may be processed using the on-device ASR model to generate the predicted output. In this example, the predicted output may correspond to a predicted target audio waveform and/or predicted audio features that is predicted to correspond to the target audio waveform portion and/or audio features of the target audio waveform portion. Put another way, the self-supervised learning technique may attempt to reconstruct the target audio waveform portion and/or the audio features thereof based on processing the prepended audio waveform portion and the appended audio waveform portion. Also, for example, in implementations where the target portion of the client data corresponds to the target representation portion of the corresponding stream of audio data, the prepended portion and the appended portion may be processed in the same or similar manner to predict the target portion. Further, the predicted output may be compared to the target portion that was masked to generate the gradient. Put another way, the actual target portion of the client data that was masked may be utilized as a supervision signal with which the predicted output may be compared. Moreover, based on the comparison, the gradient may be generated and utilized to generate the update.
In various implementations, the gradient may additionally or alternatively be generated using supervised learning. For example, again assume audio data capturing a spoken utterance of a human user of “Hey Assistant, turn on the kitchen lights” is generated via microphone(s) of a given client device of the human user. In this example, the audio data capturing the spoken utterance may correspond to the client data. Further, the audio data may be processed using an on-device ASR model stored in on-device memory of the given client device to generate ASR data, such as one or more speech hypotheses that are predicted to correspond to the spoken utterance captured in the audio data. In this example the one or more speech hypotheses may correspond to the predicted output, and a given speech hypothesis may be selected as corresponding to the spoken utterance and provided for presentation to the user as a transcription via a display of the given client device. Moreover, in this example, further assume that the user corrects one or more terms of the transcription. The corrected one or more terms of the transcription may be utilized as a supervision signal in generating the gradient for one or more of the first on-device ASR layers. For instance, the given client device may compare the one or more terms of the transcription to the one or more of the corrected terms of the transcription to generate the gradient for one or more of the first on-device ASR layers.
By using the techniques described herein, various technical advantages can be achieved. As one non-limiting example, by generating the updates based on gradients that are specific to particular portions of on-device ML models, consumption of both on-device memory and computational resources can be reduced at client devices that participate in federated learning of ML model(s). For example, the client devices store less data in generating the gradients and perform less processing in generating the gradient by only generating the gradients that are specific to the particular portion of the on-device ML models. Further, consumption of network resources can also be reduced since the updates transmitted to the remote system include less data as a function of the updates being generated based on these gradients that are specific to the particular portion of the on-device ML models. As another non-limiting example, the remote system can schedule these updates at the different client devices to enable the remote system to receive updates for the entirety of a global ML model even though no single client device generates an update for the entirety of the global ML model. As a result, the global ML models (and the on-device ML models when the global ML models are transmitted back to the client devices) have greater precision and/or recall while enabling each of the client devices that participate in the federated learning of the ML models to consume less on-device resources.
The above description is provided as an overview of some implementations of the present disclosure. Further description of those implementations, and other implementations, are described in more detail below.
Further, the client device 150 may include various on-device machine learning (ML) models stored in on-device memory of the client device 150, such as in on-device ML model(s) database 154A. In some implementations, and based on processing the client data 101, the on-device ML models can be utilized to cause certain fulfillment of the client data 101 to be performed (e.g., as described in more detail with respect to
As described herein, implementations are directed to improved efficiency of federated learning of ML model(s). Similar to traditional federated learning, on-device ML engine 132 may process, using one or more on-device ML models stored in the on-device ML model(s) database 154A, the client data 101 to generate predicted output 102. The one or more on-device ML models utilized by the on-device ML engine 132 may depend on a type of the client data 101 received at the client device 150. For example, assume the client data 101 corresponds to audio data generated via one or more microphones of the client device 150. In this example, the on-device ML engine 132 may utilize one or more audio-based ML models to process the client data 101 in generating the predicted output 102, such as an automatic speech recognition (ASR) model, a hotword detection model, a continued conversation model, a voice identification model, a voice activity detection (VAD) model, an endpointing model, and/or other audio-based ML models. Further, the predicted output 102 (and/or predicted values associated therewith, such as probabilities, log likelihoods, binary values, and/or any other predicted values) in this example may be based on the one or more audio-based ML models utilized in processing the client data 101, such as one or more predictions described with respect to the audio-based ML models of
In various implementations, to cause the on-device ML models to be updated locally at the client device 150 and/or the on-device ML models to be utilized in generating the one or more updates at the client device 150 for the counterpart global ML models, the client device 150 can cause the predicted output 102 to be provided to a gradient engine 136. The gradient engine can compare the predicted output 102 with a supervision signal 103 to generate a gradient 104. In some implementations, the supervision signal 103 may be an implicit supervision signal generated using an unsupervised learning engine 134. The unsupervised learning engine 134 can utilize one or more self-supervised learning techniques and/or semi-supervised learning techniques (e.g., as described with respect to
In some implementations, the gradient 104 (and other gradients) may be derived from a loss function used to train the ML model(s) (e.g., the on-device ML model(s) and/or the global ML model(s)), such that the gradient represents a value of that loss function (or a derivative thereof) obtained from a comparison of the supervision signal 103 to the predicted output 102. For example, when the supervision signal 103 and the predicted output 102 match, the gradient engine 136 can generate a zero gradient. Also, for example, when the supervision signal 103 and the predicted output 102 do not match, the gradient engine 136 can generate a non-zero gradient that is optionally dependent on the extent of the mismatching. The extent of the mismatching can be based on an extent of mismatching between deterministic comparisons of the supervision signal 103 to the predicted output 102. In additional or alternative implementations, the gradient 104 (and other gradients) may be derived from a loss function used to train the ML model(s), such that the gradient 104 represents a value of that loss function (or a derivative thereof) determined based on the predicted output 102 (e.g., without considering any supervision signals).
In some versions of those implementations, the gradient 104 (and other gradients) may be specific to particular ML layers of a given ML model. For example, again assume the client data 101 corresponds to audio data generated via one or more microphones of the client device 150, and again assume that the on-device ML engine 132 processes, using an on-device ASR model, the audio data to generate and select a given speech hypotheses that is predicted to correspond to a spoken utterance captured in the audio data. Further assume the on-device ASR model includes at least one or more first on-device ASR layers and one or more second on-device ASR layers. In this example, the loss function and/or the value of that loss function may be specific to the one or more first on-device ASR layers and/or specific to the one or more second on-device ASR layers. For instance, a first loss function or a first value of the first loss function may be utilized to generate the gradient 104 for the one or more first ASR layers, and a second loss function or a second value of the first loss function may be utilized to generate the gradient 104 for the one or more second ASR layers. Additionally, or alternatively, the gradient 104 for the one or more first ASR layers may be generated in this manner at the client device 150, and no gradient for the one or more second ASR layers may be generated at the client device 150. Rather, a gradient for the one or more second ASR layers may be generated at one or more additional client devices 170 in the same or similar described with respect to generating the gradient 104 at the client device 150. Each of the one or more additional client devices 170 may include corresponding instances of at least the components that are encompassed within the box of
Notably, in traditional federated learning, a plurality of client devices (e.g., the client device 150 and one or more of the additional client devices 170) each generate a corresponding gradient for a corresponding on-device ML model based on corresponding client data received a given one of the plurality of client devices. For example, each of the plurality of client devices may have a corresponding counterpart on-device ML model stored in corresponding on-device memory, and process the corresponding client data to generate gradients for the corresponding instances of the corresponding on-device ML models. Although the gradients in this example are generated at disparate client devices, the gradients are generated for the corresponding counterpart on-device ML models. Further, the remote system 160 may utilize the corresponding gradients (or corresponding updates generated based on the corresponding gradients) to update a remote-based counterpart of the corresponding counterpart on-device ML models. However, in traditional federated learning, the corresponding gradients (or the corresponding updates generated based on the corresponding gradients), are typically for the corresponding counterpart on-device ML models as a whole. In contrast, to improve efficiency in federated learning, the corresponding gradients generated may only be for a portion of the corresponding counterpart on-device ML models. As a result, consumption of computational resources and memory at each of the plurality of client devices may be reduced since the corresponding gradients generated based on the corresponding client data may only be for a portion of the corresponding counterpart on-device ML models, and consumption of network resources in transmitting the corresponding gradients to the remote system 160 may be reduced.
Accordingly, in various implementations, the client device 150 (and one or more of the additional client devices 170) may include on-device scheduling engine 138 to ensure the corresponding gradients are generated for disparate portions of the corresponding counterpart on-device ML models at each of the plurality of client devices. The on-device scheduling signal 138 may generate a scheduling signal 105 that is provided to the gradient engine 136 to ensure the client device 150 is generating the gradient for the appropriate on-device ML layers of the corresponding on-device ML model. For example, the on-device scheduling engine 138 of the client device 150 may instruct the gradient engine 136 to generate the gradient 104 for one or more first on-device ML layers of the corresponding counterpart on-device ML model via the scheduling signal 105. Further, a corresponding on-device scheduling engine of one or more of the additional client devices 170 may instruct a corresponding gradient engine to generate a corresponding gradient for one or more second on-device ML layers of the corresponding counterpart on-device ML model via a corresponding scheduling signal. The scheduling signal 105 for the client device 150 and the corresponding scheduling signals for one or more of the additional client devices 170 may be received at each of the client devices from a remote scheduling engine 146 of the remote system (e.g., as described with respect to
In some implementations, the gradient engine 136 can cause the gradient 104 to be stored in on-device memory of the client device 150 along with other gradients generated at the client device 150 until one or more client device conditions are satisfied (e.g., described with respect to
In additional or alternative implementations, the ML update engine 140 can cause the corresponding counterpart on-device ML model to be updated based on the gradient 104. Continuing with the above example, again assume the on-device scheduling engine 138 instructs the gradient engine 136 to generate the gradient 104 for one or more first on-device ML layers of the corresponding counterpart on-device ML model via the scheduling signal 105. In this example, the ML update engine 140 can cause the gradient 104 to be backpropagated across the one or more first on-device ML layers to update one or more corresponding first on-device weights of the one or more first on-device ML layers, and while the one or more second on-device ML layers (and/or other on-device ML layers) are fixed (e.g., not updated), thereby resulting in one or more updated first on-device ML layers. Further, the same process may be repeated at the given one of the additional client devices 170 to update one or more corresponding second on-device weights of the one or more second on-device ML layers, and while the one or more first on-device ML layers (and/or other on-device ML layers) are fixed, thereby resulting in one or more updated second on-device ML layers.
In some versions of these additional or alternative implementations, the ML update 106 can include the one or more updated first on-device ML layers and an indication of one or more global ML layers of a global ML model that should be replaced in the remote memory of the remote system 160 with the one or more updated first on-device ML layers. Continuing with the above example, the ML update 106 may include an indication that the one or more first global ML layers of the corresponding counterpart global ML model should be replaced with the one or more updated first on-device ML layers. Similarly, the additional ML update 107 may include an indication that the one or more second global ML layers of the corresponding counterpart global ML model should be replaced with the one or more updated second on-device ML layers.
In other versions of these additional or alternative implementations, the ML update 106 can include the one or more updated first on-device weights for the one or more first updated on-device ML layers and an indication of one or more global weights a global ML model that should be replaced in the remote memory of the remote system 160 with the one or more updated first on-device weights. Continuing with the above example, the ML update 106 may include an indication that one or more first global weights of the one or more first global ML layers of the corresponding counterpart global ML model should be replaced with the one or more updated first on-device weights for the one or more first updated on-device ML layers. Similarly, the additional ML update 107 may include an indication that one or more second global weights of the one or more second global ML layers of the corresponding counterpart global ML model should be replaced with one or more updated second on-device weights for the one or more second updated on-device ML layers.
Put another way, in traditional federated learning, the corresponding counterpart global ML model may be updated in a distributed manner in that the corresponding counterpart global ML model is updated based on gradients (e.g., the gradient 104 and/or one or more additional gradients) or updates (e.g., the ML update 106 and/or one or more of the additional ML updates 107) generated at a plurality of client devices. However, a considerable amount of computational and network resources may be consumed at each of the plurality of client devices in generating these gradients or updates. Accordingly, by utilizing techniques described herein, both computational and network resources that would otherwise be consumed at each of the plurality of client devices may be conserved by enabling each of the client devices to generate and transmit less data compared to traditional federated learning.
The remote update engine 142 can utilize at least the ML update 106 received from the client device 150 and/or one or more of the additional ML updates 107 to update one or more of the global ML models stored in the global ML model(s) database 154B. In some implementations, the ML update 106 received from the client device 150 and/or one or more of the additional ML updates 107 may be stored in an update(s) database 180 when they are received at the remote system 160. In some versions of those implementations, the remote update engine 142 may obtain the updates to update one or more of the global ML models when one or more remote system conditions are satisfied (e.g., as described with respect to
Notably, the remote update engine 142 may update the corresponding global ML models in various manners based on a type of the ML update 106 is received from the client device 150 and/or as one or more of the additional ML updates 107 is received from one or more of the additional client devices 170. For example, the remote update engine 142 can identify a particular global ML model, from among the global ML models stored in the global ML model(s) database 154B, to update global layers and/or global weights thereof. In some implementations, the remote update engine 142 can identify the particular global ML model based on the indication of the type of global ML model to be updated based on the ML update 106 and/or one or more of the additional ML updates 107. For example, if a plurality of updates for an ASR model are received and/or stored in the update(s) database 180, the remote update engine 142 can identify the corresponding counterpart global ASR model to be updated. Further, how the remote update engine 142 updates the corresponding counterpart global ASR model may be based on content of the ML update 106 and/or one or more of the additional ML updates 107. For instance, if a given update includes a gradient and indicates that is it for one or more first on-device ASR layers of an on-device ASR model, then the remote update engine 142 may backpropagate the gradient across one or more corresponding counterpart first global ASR layers of a corresponding counterpart global ASR model while at least one or more second global ASR layers of the corresponding counterpart global ASR model are fixed. Also, for instance, if a given update includes one or more first updated on-device ASR layers of an on-device ASR model and/or one or more updated first on-device weights for the one or more first updated on-device ASR layers of the on-device ASR model, then the remote update engine 142 may replace one or more global ASR layers of a corresponding counterpart ASR model with the one or more first updated on-device ASR layers and/or replace one or more first global weights for the one or more global ASR layers with the one or more updated first on-device weights for the one or more first updated on-device ASR layers of the on-device ASR model. In various implementations, updates stored in the update(s) database 180 may include updates of different types. In these implementations, the remote update engine 142 may utilize one or more criteria to determine how and/or when to utilize the different types of updates. The one or more criteria may indicate which types of updates should be prioritized over other types of updates. As one non-limiting example, the one or more criteria may indicate that updates that replace layers and/or weights should be prioritized over updates that include gradients, but it should be understood that other criteria are contemplated herein. Further, in some implementations, the remote update engine 142 can identify a single global ML model to be updated at a given time instance, whereas in other implementations, the remote update engine 142 can identify multiple global ML models to be updated, in parallel, at the given time instance.
In some implementations, the remote update engine 142 can assign the updates to specific iterations of updating of one or more of the global ML models based on one or more criteria. The one or more criteria can include, for example, the types of updates available to the remote update engine 142, a threshold quantity of updates available to the remote update engine 142 (and optionally for the disparate portions of the one or more global ML models), a threshold duration of time of required to perform the updates, and/or other criteria. In particular, the remote update engine 142 can identify multiple sets or subsets of updates generated by the client devices 150 and/or 170. Further, the remote update engine 142 can update one or more of the global ML models based on these sets or subsets of the updates. In some further versions of those implementations, a quantity of updates in the sets or subsets may be the same or vary. In yet further versions of those implementations, each of the sets or subsets of updates can optionally include updates from at least two unique client devices. In other implementations, the remote update engine 142 may utilize the updates to update one or more of the global ML models in a first in, first out (FIFO) manner without assigning any update to a specific iteration of updating of one or more of the global ML models.
Update distribution engine 144 can transmit one or more of the updated global ML models 108 (e.g., including the updated global ML model itself, updated layers thereof, and/or updated weights thereof) to the client device 150 and/or one or more of the additional client devices 170. In some implementations, the update distribution engine 144 can transmit one or more of the updated global ML models 108 in response to one or more remote system conditions being satisfied at the remote system 160 and/or one or more client device conditions being satisfied at the client device 150 and/or one or more of the additional client device 170 (e.g., as described with respect to
Turning now to
The client device 250 in
One or more cloud-based automated assistant components 270 can optionally be implemented on one or more computing systems (collectively referred to as a “cloud” computing system) that are communicatively coupled to client device 250 via one or more networks as indicated generally by 299. The cloud-based automated assistant components 270 can be implemented, for example, via a cluster of high-performance servers. In various implementations, an instance of the automated assistant client 240, by way of its interactions with one or more of the cloud-based automated assistant components 270, may form what appears to be, from a user's perspective, a logical instance of an automated assistant as indicated generally by 295 with which the user may engage in a human-to-computer interactions (e.g., typed interactions, spoken interactions, gesture-based interactions, and/or touch-based interactions).
The client device 250 can be, for example: a desktop computing device, a laptop computing device, a tablet computing device, a mobile phone computing device, a computing device of a vehicle of the user (e.g., an in-vehicle communications system, an in-vehicle entertainment system, an in-vehicle navigation system), a standalone interactive speaker, a smart appliance such as a smart television (or a standard television equipped with a networked dongle with automated assistant capabilities), and/or a wearable apparatus that includes a computing device (e.g., a watch of the user having a computing device, glasses of the user having a computing device, a virtual or augmented reality computing device). Additional and/or alternative client devices may be provided.
The one or more vision components 213 can take various forms, such as monographic cameras, stereographic cameras, a LIDAR component (or other laser-based component(s)), a radar component, etc. The one or more vision components 213 may be used, e.g., by the visual capture engine 218, to capture image data corresponding to vision frames (e.g., image frames, laser-based vision frames) of an environment in which the client device 250 is deployed. In some implementations, such vision frame(s) can be utilized to determine whether a user is present near the client device 250 and/or a distance of a given user of the client device 250 (e.g., the user's face) relative to the client device 250. Such determination(s) can be utilized, for example, in determining whether to activate the various on-device ML engines depicted in
As described herein, such audio data and/or other non-audio sensor data such as vision data and textual data (also collectively referred to herein as client data) can be processed by the various engines depicted in
As some non-limiting example, the respective hotword detection engines 222, 272 can utilize respective hotword detection models 222A, 272A to predict whether audio data includes one or more particular words or phrases to invoke the automated assistant 295 (e.g., “Ok Assistant”, “Hey Assistant”, “What is the weather Assistant?”, etc.) or certain functions of the automated assistant 295 (e.g., “Stop” when an alarm is sounding or music is playing, “Next” when viewing images or listening to music, etc.); the respective hotword free invocation engines 224, 274 can utilize respective hotword free invocation model 224A, 274A to predict whether non-audio sensor data (e.g., vision data) includes a gesture or signal to invoke the automated assistant 295 (e.g., based on a gaze of the user and optionally further based on mouth movement of the user); the respective continued conversation engines 226, 276 can utilize a respective continued conversation model 226A, 276A to predict whether further audio data is directed to the automated assistant 295 (e.g., or directed to an additional user in the environment of the client device 250); the respective ASR engines 228, 278 can utilize a respective ASR model 228A, 278A to generate recognized text, or predict phoneme(s) and/or token(s) that correspond to audio data detected at the client device 250 and generate the recognized text based on the phoneme(s) and/or token(s); the respective object detection engines 230, 280 can utilize a respective object detection model 230A, 280A to predict object location(s) included in vision data captured at the client device 250; the respective object classification engines 232, 282 can utilize a respective object classification model 232A, 282A to predict object classification(s) of object(s) included vision data captured at the client device 250; the respective voice identification engines 234, 284 can utilize a respective voice identification model 234, 284A to predict whether audio data captures a spoken utterance of one or more users of the client device 250 (e.g., by generating a speaker embedding, or other representation, that can be compared to a corresponding actual embeddings for one or more of the user of the client device 250); and the respective face identification engines 236, 286 can utilize a respective face identification model 236A, 286A to predict whether image data captures one or more of the users in an environment of the client device 250 (e.g., by generating an image embedding, or other representation, that can be compared to a corresponding image embeddings for one or more of the user of the client device 250). Each of these various ML models (and other ML models) can be updated in the manner described herein.
In some implementations, the client device 250 and one or more of the cloud-based automated assistant components 270 may further include natural language understanding (NLU) engines 238, 294 and fulfillment engine 240, 296, respectively. The NLU engines 238, 294 may perform natural language understanding, utilizing respective NLU models 238A, 294A, on recognized text, predicted phoneme(s), and/or predicted token(s) generated by the ASR engines 228, 278 to generate NLU data. The NLU data can include, for example, intent(s) that correspond to the spoken utterance and optionally slot value(s) for parameter(s) for the intent(s). Further, the client device 250 and one or more of the cloud-based automated assistant components 270 may further include fulfillment engines 240, 296, respectively. The fulfillment engines 240, 296 can generate fulfillment data utilizing respective fulfillment models or rules 240A, 296A, and based on processing the NLU data. This fulfillment data can define certain fulfillment that is responsive to user input (e.g., spoken utterances, typed input, touch input, gesture input, and/or any other type of user input) provided by a user of the client device 250. The certain fulfillment can include interaction(s) to perform with locally installed application(s) based on the user input, command(s) to transmit to Internet-of-things (IoT) device(s) (directly or via corresponding remote system(s)) based on the user input, and/or other resolution action(s) to perform based on the user input. The fulfillment data is then provided for local and/or remote performance/execution of the determined action(s) to cause the certain fulfillment of the user input to be performed. Execution can include, for example, rendering local and/or remote responses (e.g., visually and/or audibly rendering (optionally utilizing an on-device TTS module)), interacting with locally installed applications, transmitting command(s) to IoT device(s), and/or other action(s). In other implementations, the NLU engines 238, 294 and the fulfillment engines 240, 296 may be omitted, and the ASR engines 228, 278 can generate the fulfillment data directly based on the user input. For example, assume one or more of the ASR engines 228, 278 process, using the respective ASR model 228A, 278A, a spoken utterance of “turn on the lights.” In this example, the ASR engines 228, 278 can generate a semantic output that is then transmitted to a software application associated with the lights and/or directly to the lights that indicates that they should be turned on, and optionally a transcription of the spoken utterance to be visually rendered for presentation to a user that provided the spoken utterance.
Notably, the cloud-based automated assistant component(s) 270 include remote-based counterparts to the engines and models described herein with respect to
Turning now to
At block 352, the system receives, at a client device and from a user of the client device, client data. The client data may be one or more of various types. The various types of client data can include, for example, audio data generated via microphone(s) of the client device, vision data generated via vision component(s), textual data or touch data generated via user interface input component(s) of the client device (e.g., via a touchscreen display, a mouse and keyboard, and/or other user interface components), and/or any other client data. In some implementations, the client data may be directed to automated assistant (e.g., to cause visual and/or audible content to be provided, to cause an IoT device to be controlled, etc.), whereas in other implementations, the client may be provided independent of an automated assistant (e.g., dictating an email or note without interacting with an automated assistant, etc.).
At block 354, the system processes, using an on-device ML model stored in on-device memory of the client device, the client data to generate predicted output, the on-device ML model including at least one or more first on-device ML layers and one or more second on-device ML layers. The on-device ML model utilized to generate the predicted output may be based on the type of client data received at block 352. For example, in implementations where the client data received at block 352 is audio data, the system may utilize an on-device audio-based ML model to process the client data to generate the predicted. Further, the predicted output may be based on a type of the audio-based ML model utilized to process the client data. For instance, if the on-device audio-based ML model is an on-device ASR model, then the predicted output may be a predicted speech hypothesis that is predicted to correspond to a spoken utterance (if any) captured in the audio data. Also, for instance, if the on-device audio-based ML model is an on-device hotword detection model, then the predicted output may include an indication of whether the audio data is predicted to include a particular word or phrase that invokes or otherwise controls an automated assistant.
At block 356, the system generates, using unsupervised learning, a gradient based on the predicted output. Generating the gradient using unsupervised learning is described in more detail herein (e.g., with respect to
The system returns to block 352 to receive additional client data and performs a subsequent iteration of the operations of blocks 352-358. In some implementations, and during the subsequent iteration of the operations of blocks 352-358, the system may generate an additional first update for the one or more first on-device ML layers of the on-device ML model. In additional or alternative implementations, and during the subsequent iteration of the operations of blocks 352-358, the system may generate a second update for the one or more second on-device ML layers of the on-device ML model. For example, the client device may be utilized to generate first updates for only the one or more first on-device ML layers of the on-device ML model, or the client device may alternate between generating first updates and second updates for the disparate on-device ML layers of the on-device ML model. In some implementations, a scheduling engine (e.g., the on-device scheduling engine 138 of
At block 360, the system determines whether one or more client device conditions are satisfied for transmitting the first update (and/or any other updates generated during one or more subsequent iterations of the operations of blocks 352-358) to a remote system. The one or more client device conditions may include, for example, a particular time of day, a particular day of week, whether the client device is charging, whether the client device has at least a threshold state of charge, whether a temperature of the client device is less than a temperature threshold, and/or whether the client device is being held by a user. If, at an iteration of block 360, the system determines that the one or more client device conditions are not satisfied, then the system continues monitoring for satisfaction of the one or more client device conditions at block 360. Notably, while the system continues monitoring for satisfaction of the one or more client device conditions at block 360, the system may continue generating updates via subsequent iterations of the operations of blocks 352-358. If, at an iteration of block 360, the system determines that the one or more client device conditions are satisfied, then the system proceeds to block 362.
At block 362, the system transmits the first update for the one or more first on-device ML layers to a remote system to cause the remote system to update one or more first global ML layers of a global ML model, that is a remote-based counterpart of the on-device ML model, based on the first update and while one or more second global ML layers of the global ML model are fixed. Notably, the global ML model is a remote-based counterpart of the on-device ML model and includes the one or more first global ML layers and the one or more second global ML layers. Accordingly, the one or more first global ML layers may be a remote-based counterpart of the one or more first on-device ML layers. However, the first update described with respect to
At block 364, the system determines whether one or more remote system conditions are satisfied for transmitting the updated global ML model (e.g., the updated global ML model itself, one or more updated global ML layers of the updated global ML model, or one or more updated global weights of the updated global ML model) to the client device. The one or more remote system conditions may include, for example, a particular time of day, a particular day of week, whether a threshold quantity of updates have been utilized to update the updated global ML model, and/or whether performance of the updated global ML model satisfies a performance threshold. If, at an iteration of block 364, the system determines that the one or more client device conditions are not satisfied, then the system continues monitoring for satisfaction of the one or more remote system conditions at block 364. Notably, while the system continues monitoring for satisfaction of the one or more remote system conditions at block 364, the system may continue updating the global ML model based on updates received from the client device and/or one or more additional client devices. If, at an iteration of block 364, the system determines that the one or more remote system conditions are satisfied, then the system proceeds to block 366.
At block 366, the system receives, at the client device, the updated global ML model (e.g., including the updated global ML model itself, one or more updated global ML layers thereof, and/or one or more updated global weights thereof). Although not depicted for the sake of brevity, the client device may only receive the updated global model in response to determining one or more client device conditions are satisfied (e.g., as described with respect to the operations of block 360). In response to receiving the updated global ML model, the client device may replace, in the on-device memory of the client device, the on-device ML model with the updated global ML model, resulting in an updated on-device ML model. However, the client device (and one or more additional client devices) may continue generating additional updates for the updated on-device ML model, and an instance of the updated global ML model may be maintained at the remote system to be further updated based on the additional updates.
Although the method 300 of
Turning now to
At block 452, the system identifies a target portion of the client data (e.g., the client data received at an iteration of the operations of block 352 of
At block 456, the system processes, using the on-device ML model, the prepended portion of the client data and the appended portion of the client data to generate predicted outputs, the predicted output predicting the target portion of the client data. For example, in implementations where the target portion of the client data corresponds to the target audio waveform portion of the audio data, further assume the audio data captures a spoken utterance. In this example, the system can process, using an on-device ASR model, the prepended audio waveform portion and the appended audio waveform portion to generate the predicted output. In this example, the predicted output may correspond to a predicted target audio waveform and/or one or more predicted features of the predicted target audio waveform (e.g., a predicted amplitude, a predicted wavelength, a predicted phase, a predicted period, and/or other features) that are predicted to correspond to the target audio waveform portion. Additionally, or alternatively, the predicted output may more generally correspond to predicted target audio data and/or one or more predicted features of the predicted audio data (e.g., predicted MFCCs, predicted melbank features, and/or other features) that are predicted to correspond to the target audio data. Put another way, the system may attempt to reconstruct the target audio waveform portion and/or features of the audio data based on processing the prepended audio waveform portion and the appended audio waveform portion.
Notably, in implementations where the on-device ASR model is utilized to reconstruct the target audio waveform portion, the on-device ASR model may be a multilingual on-device ASR model since a primary language of a given client device that generated the audio data may be irrelevant to reconstructing the target audio waveform portion. However, in various implementations, the system may additionally, or alternatively, process the prepended audio waveform portion and the appended audio waveform portion using one or more on-device ASR models that are specific to the primary language of the given client device that generated audio data. Also, for example, in implementations where the target portion of the client data corresponds to the target representation portion of the audio data, further assume the audio data captures a spoken utterance. In this example, the system can process, using an on-device ASR model and/or an on-device NLU model, the prepended representation portion and the appended representation portion to generate the predicted output. In this example, the predicted output may correspond to a predicted target representation that is predicted to correspond to the target representation portion. Put another way, the system may attempt to reconstruct the target representation portion based on processing the prepended representation portion and the appended representation portion.
At block 458, the system compares the predicted output to the target portion of the client data. At block 460, the system generates, based on comparing the predicted output to the target portion of the client data, the gradient. Put another way, the system may utilize the actual target portion of the client data that was masked at block 454 as a supervision signal with which the predicted output that is predicted to correspond to the target portion may be compared at block 458. Further, based on the comparison, the system may generate the gradient. The system may perform additional iterations of the method 400 of
Although the method 400 of
Turning now to
At block 552, the system compresses, at the client device (e.g., the client data received at an iteration of the operations of block 352 of
Further, at block 552A, the system compresses, at an additional client device and prior to generating an additional gradient (e.g., based on additional client data received at the additional client device), at least the second on-device ML layer and a third on-device ML layer of the on-device ML model into the one or more second ML layers. The system may compress the second on-device ML layer and the third on-device ML layer of the on-device ML model into the one or more second ML layers in the same or similar manner, but based on a scheduling signal generated by an on-device scheduling engine of the additional client device and/or received from a remote scheduling engine (e.g., the remote scheduling engine 146 of
At block 554, the system generates, at the client device, the first update (e.g., the first update generated at the first iteration of the operations of block 358 of
At block 556, the system causes a first global ML layer of the global ML model (e.g., the global ML model to be updated in
Turning now to
At block 652, the system schedules one or more updates for the global ML model (e.g., the global ML model to be updated in
In some implementations, the remote scheduling engine (e.g., the remote scheduling engine 146 of
At block 654, the system causes one or more Nth global ML layers of the global ML model (e.g., the global ML model to be updated in
Turning now to
At block 752, the system receives, at a remote system and from one or more client devices, one or more updates for a global ML model stored remotely in remote memory of the remote system, the global ML model including a plurality of global ML layers, the plurality of global ML layers including at least one or more first global ML layers and one or more second global ML layers. The one or more updates for the global ML model may be generated at one or more of the client devices in any manner described herein (e.g., with respect to
At block 754, the system determines whether one or more update conditions are satisfied for updating the global ML model. The one or more update conditions for updating the global ML model may include, for example, a particular time of day, a particular day of week, whether a threshold quantity of updates are available to update global ML model, and/or whether a threshold quantity of disparate updates is available to update global ML model. If, at an iteration of block 754, the system determines that the one or more update conditions are not satisfied, the system returns to block 752 to receive one or more additional updates from one or more of the client devices. If, at an iteration of block 754, the system determines that the one or more update conditions are satisfied, the system proceeds to block 756.
At block 756, the system causes, based on the one or more updates received from one or more of the client devices, the global ML model to be updated to generate an updated global ML model. For example, at block 756A, the system causes only the one or more first global ML layers to be updated based on a first update received from a client device of the one or more client devices while at least one of the one or more second global ML layers are fixed. Also, for example, at block 756B, the system causes one or more second global ML layers to be updated based on a second update received from an additional client device of the one or more client devices while at least one of the one or more first global ML layers are fixed.
At block 758, the system determines whether one or more client device conditions are satisfied at one or more of the client devices. The one or more client device conditions may include, for example, a particular time of day, a particular day of week, whether the client device is charging, whether the client device has at least a threshold state of charge, whether a temperature of the client device is less than a temperature threshold, and/or whether the client device is being held by a user. If, at an iteration of block 758, the system determines that the one or more client device conditions are not satisfied, then the system continues monitoring for satisfaction of the one or more client device conditions at block 758. Notably, while the system continues monitoring for satisfaction of the one or more client device conditions at block 758, the system may continue updating the global ML model via subsequent iterations of the operations of blocks 752-756. If, at an iteration of block 758, the system determines that the one or more client device conditions are satisfied, then the system proceeds to block 760.
At block 760, the system transmits, to one or more of the client devices, the updated global ML model (e.g., the updated global ML model itself, one or more updated global ML layers of the updated global ML model, or one or more updated global weights of the updated global ML model). Although not depicted for the sake of brevity, one or more of the client devices may only receive the updated global model in response to determining one or more client device conditions are satisfied (e.g., as described with respect to the operations of block 360). In response to transmitting the updated global ML model, the client device may cause the on-device ML model to be replaced, in the on-device memory of the client device, the on-device ML model with the updated global ML model, resulting in an updated on-device ML model. However, the client device (and one or more additional client devices) may continue generating additional updates for the updated on-device ML model, and an instance of the updated global ML model may be maintained at the remote system to be further updated based on the additional updates.
Turning now to
Computing device 810 typically includes at least one processor 814 which communicates with a number of peripheral devices via bus subsystem 812. These peripheral devices may include a storage subsystem 824, including, for example, a memory subsystem 825 and a file storage subsystem 826, user interface output devices 820, user interface input devices 822, and a network interface subsystem 816. The input and output devices allow user interaction with computing device 810. Network interface subsystem 816 provides an interface to outside networks and is coupled to corresponding interface devices in other computing devices.
User interface input devices 822 may include a keyboard, pointing devices such as a mouse, trackball, touchpad, or graphics tablet, a scanner, a touchscreen incorporated into the display, audio input devices such as voice recognition systems, microphones, and/or other types of input devices. In general, use of the term “input device” is intended to include all possible types of devices and ways to input information into computing device 810 or onto a communication network.
User interface output devices 820 may include a display subsystem, a printer, a fax machine, or non-visual displays such as audio output devices. The display subsystem may include a cathode ray tube (CRT), a flat-panel device such as a liquid crystal display (LCD), a projection device, or some other mechanism for creating a visible image. The display subsystem may also provide non-visual display such as via audio output devices. In general, use of the term “output device” is intended to include all possible types of devices and ways to output information from computing device 810 to the user or to another machine or computing device.
Storage subsystem 824 stores programming and data constructs that provide the functionality of some or all of the modules described herein. For example, the storage subsystem 824 may include the logic to perform selected aspects of the methods disclosed herein, as well as to implement various components depicted in
These software modules are generally executed by processor 814 alone or in combination with other processors. Memory 825 used in the storage subsystem 824 can include a number of memories including a main random access memory (RAM) 830 for storage of instructions and data during program execution and a read only memory (ROM) 832 in which fixed instructions are stored. A file storage subsystem 826 can provide persistent storage for program and data files, and may include a hard disk drive, a floppy disk drive along with associated removable media, a CD-ROM drive, an optical drive, or removable media cartridges. The modules implementing the functionality of certain implementations may be stored by file storage subsystem 826 in the storage subsystem 824, or in other machines accessible by the processor(s) 814.
Bus subsystem 812 provides a mechanism for letting the various components and subsystems of computing device 810 communicate with each other as intended. Although bus subsystem 812 is shown schematically as a single bus, alternative implementations of the bus subsystem may use multiple busses.
Computing device 810 can be of varying types including a workstation, server, computing cluster, blade server, server farm, or any other data processing system or computing device. Due to the ever-changing nature of computers and networks, the description of computing device 810 depicted in
In situations in which the systems described herein collect or otherwise monitor personal information about users, or may make use of personal and/or monitored information), the users may be provided with an opportunity to control whether programs or features collect user information (e.g., information about a user's social network, social actions or activities, profession, a user's preferences, or a user's current geographic location), or to control whether and/or how to receive content from the content server that may be more relevant to the user. Also, certain data may be treated in one or more ways before it is stored or used, so that personal identifiable information is removed. For example, a user's identity may be treated so that no personal identifiable information can be determined for the user, or a user's geographic location may be generalized where geographic location information is obtained (such as to a city, ZIP code, or state level), so that a particular geographic location of a user cannot be determined. Thus, the user may have control over how information is collected about the user and/or used.
In some implementations, a method performed by one or more processors of a client device is provided and includes receiving, from a user of the client device, client data, the client data being generated locally at the client device the client device; and processing, using an on-device machine learning (ML) model stored locally in on-device memory of the client device, the client data to generate predicted output. The on-device machine learning model includes a plurality of on-device ML layers, and the plurality of on-device ML layers include at least one or more first on-device ML layers and one or more second on-device ML layers. The method further includes generating, using unsupervised learning, a gradient based on the predicted output; generating, based on the gradient, a first update for the one or more first on-device ML layers of the on-device ML model stored locally in the on-device memory of the client device; and transmitting the first update to a remote system. Transmitting the first update to the remote system causes the remote system to update a global ML model stored remotely in remote memory of the remote system, the global ML model includes at least one or more first global ML layers and one or more second global ML layers. Further, causing the remote system to update the global ML model includes causing the one or more first global ML layers to be updated based on the first update while one or more of the second global ML layers are fixed.
These and other implementations of the technology can include one or more of the following features.
In some implementations, the first update transmitted to the remote system may include the gradient and an indication of the one or more first global ML layers to be updated based on the first update, and the one or more first global ML layers of the global ML model stored remotely at the remote system may correspond to the one or more first on-device ML layers of the on-device ML model stored locally at the client device. In some versions of those implementations, causing the one or more first global ML layers to be updated based on the first update while one or more of the second global ML layers are fixed may include causing, based on the gradient and based on the indication of the one or more first global ML layers to be updated based on the first update, the one or more first global ML layers to be updated based on the gradient to generate one or more updated first global ML layers without updating the one or more second global ML layers, the one or more updated first global ML layers including one or more updated first global weights for the one or more updated first global ML layers.
In some implementations, generating the first update for the one or more first on-device ML layers may include causing the one or more first on-device ML layers to be updated based on the gradient to generate one or more updated first on-device ML layers without updating the one or more second on-device ML layers, the one or more updated first on-device ML layers including one or more updated first on-device weights for the one or more updated first on-device ML layers.
In some versions of those implementations, the first update transmitted to the remote system may include the one or more updated first on-device ML layers and an indication of the one or more first global ML layers to be updated based on the first update, and wherein the one or more first global ML layers of the global ML model stored remotely at the remote system correspond to the one or more first on-device ML layers of the on-device ML model stored locally at the client device. In some further versions of those implementations, causing the one or more first global ML layers to be updated based on the first update while one or more of the second global ML layers are fixed may include causing, based on the one or more updated first on-device ML layers and based on the indication of the one or more first global ML layers to be updated based on the first update, the one or more first global ML layers to be replaced in the remote memory with the one or more updated first on-device ML layers without replacing the one or more second global ML layers.
In additional or alternative versions of those implementations, the first update transmitted to the remote system may include the one or more updated first on-device weights for the one or more updated first on-device ML layers and an indication of the one or more first global ML layers to be updated based on the first update, and the one or more first global ML layers of the global ML model stored remotely at the remote system may correspond to the one or more first on-device ML layers of the on-device ML model stored locally at the client device. In some further versions of those implementations, causing the one or more first global ML layers to be updated based on the first update while one or more of the second global ML layers are fixed may include causing, based on the one or more updated first on-device weights for the one or more updated first on-device ML layers and based on the indication of the one or more first global ML layers to be updated based on the first update, one or more first global weights for the one or more first global ML layers to be replaced in the remote memory of the remote system with the one or more updated first on-device weights for the one or more updated first on-device ML layers without replacing one or more second global weights for the one or more second global ML layers.
In some implementations, causing the remote system to update the global ML model may further include causing the one or more second global ML layers to be updated based on a second update while one or more of the first global ML layers are fixed, and the second update may be transmitted to the remote system from an additional client device that is in addition to the client device utilized to generate the first update. In some versions of those implementations, the method may further include receiving, at the client device and from the remote system, an updated global ML model, the updated global ML model including at least the one or more updated first global ML layers and the one or more updated second global ML layers; and replacing, in the on-device memory of the client device, the on-device ML model with the updated global ML model.
In some further versions of those implementations, receiving the updated global ML model may be in response to determining one or more remote system conditions are satisfied at the remote system. The one or more remote system conditions may include one or more of: a particular time of day, a particular day of week, whether a threshold quantity of updates have been utilized to update the updated global ML model, or whether performance of the updated global ML model satisfies a performance threshold. In yet further versions of those implementations, receiving the updated global ML model may be further in response to determining one or more client device conditions are satisfied at the client device. The one or more client device conditions may include one or more of: a particular time of day, a particular day of week, whether the client device is charging, whether the client device has at least a threshold state of charge, whether a temperature of the client device is less than a temperature threshold, or whether the client device is being held by a user.
In additional or alternative further versions of those implementations, the updated global ML model received at the client device and from the remote system may include one or more of: the updated global ML model that, when received, causes the client device to replace, in the on-device memory of the client device, the on-device ML model with the updated global ML model; the one or more updated first global ML layers that, when received, causes the client device to replace, in the on-device memory of the client device, the one or more first on-device ML layers with the one or more updated first global ML layers; the one or more updated second global ML layers that, when received, causes the client device to replace, in the on-device memory of the client device, the one or more second on-device ML layers with the one or more updated second global ML layers; one or more updated first global weights for the one or more updated first global ML layers that, when received, causes the client device to replace, in the on-device memory of the client device, one or more first local weights for the one or more first on-device ML layers with the one or more updated first global weights; or one or more updated second global weights for the one or more updated second global ML layers that, when received, causes the client device to replace, in the on-device memory of the client device, one or more second local weights for the one or more second on-device ML layers with the one or more updated second global weights.
In some implementations, the method may further include identifying a target portion of the client data, the target portion of the client data being subsequent to a prepended portion of the client data that is received prior to the target portion, and the target portion of the client data being prior to an appended portion of the client data that is received subsequent to the target portion; and masking the target portion of the client data. Processing the client data using the on-device ML model to generate the predicted output may include processing the prepended portion of the client data and the appended portion of the client data to generate one or more of a predicted target portion of the client data that is predicted to correspond to the target portion of the client data. In some versions of those implementations, generating the gradient based on the predicted output using unsupervised learning may include comparing the predicted target portion of the client data to the target portion of the client data; and generating the gradient based on comparing the predicted target portion to the target portion.
In some implementations, the one or more first on-device ML layers may include at least a first on-device ML layer and a second on-device ML layer, the method may further include, prior to generating the gradient based on the predicted output: compressing the first on-device ML layer and the second on-device ML layer into the one or more first on-device ML layers, wherein the first update for the one or more on-device ML layers is a first shared update for the first on-device ML layer and the second on-device ML layer. In some versions of those implementations, causing the one or more first global ML layers to be updated based on the first update while one or more of the second global ML layers are fixed may include utilizing the first update to update a first global ML layer corresponding to the first on-device ML layer; and utilizing the first update to update a second global ML layer corresponding to the second on-device ML layer. In additional or alternative versions of those implementations, the one or more second on-device ML layers may include at least the second on-device ML layer and a third on-device ML layer, the second on-device ML layer and the third on-device ML layer may be compressed at the additional client device into the one or more second on-device ML layers at the additional client device, and a second update generated locally at the additional client device may be a shared second update for the second on-device ML layer and the third on-device ML layer. In additional or alternative versions of those implementations, causing the remote system to update the global ML model may further include causing the one or more second global ML layers to be updated based on the second update while one or more of the first global ML layers are fixed, and causing the one or more of the second global ML layers to be updated based on the second update while one or more of the first global ML layers are fixed may include utilizing the second update to update the second global ML layer corresponding to the second on-device ML layer; and utilizing the second update to update a third global ML layer corresponding to the third on-device ML layer.
In some implementations, a method performed by one or more processors of a client device is provided and includes receiving, from a client device of a user and at the remote system, a first update for a global machine learning (ML) model stored remotely at the remote system. The global ML model includes a plurality of global ML layers, and the first update for the global ML model is only for one or more first global ML layers, of the plurality of global ML layers, of the global ML model. The method further includes receiving, from an additional client device of an additional user and at the remote system, a second update for the global ML model stored remotely in remote memory of the remote system. The second update for the global ML model is only for one or more second global ML layers, of the plurality of global ML layers, of the global ML model, and the one or more second global ML layers of the global ML model are distinct from the one or more first global ML layers of the global ML model. The method further includes causing, based on at least the first update received from the client device and the second update received from the additional client device, the global ML model to be updated to generate an updated global ML model; and in response to determining one or more conditions are satisfied: transmitting the updated global ML model to one or more of: the client device, the additional client device, or one or more further additional client devices.
These and other implementations of the technology can include one or more of the following features.
In some implementations, causing the global ML model to be updated to generate the updated global ML model may include causing only the one or more first global ML layers to be updated based on the first update while the one or more second global ML layers are fixed to generate one or more updated first global ML layers; causing only the one or more second global ML layers to be updated based on the second update while the one or more first global ML layers are fixed to generate one or more updated second global ML layers; and combining the one or more updated first global ML layers and the one or more second global ML layers to generate the updated global ML model.
In some versions of those implementations, causing only the one or more first global ML layers to be updated based on the first update while the one or more second global ML layers are fixed and causing only the one or more second global ML layers to be updated based on the second update while the one or more first global ML layers are fixed may occur during a same iteration of updating the global ML model. In additional or alternative versions of those implementations, causing only the one or more first global ML layers to be updated based on the first update while the one or more second global ML layers are fixed may occur during a first iteration of updating the global ML model. Causing only the one or more second global ML layers to be updated based on the second update while the one or more first global ML layers are fixed may occur during a second iteration of updating the global ML model. The second iteration of updating the global ML model may be distinct from the first iteration of updating the global ML model.
In some implementations, causing the global ML model to be updated to generate an updated global ML model based on at least the first update received from the client device and the second update received from the additional client device may be in response to determining one or more update conditions are satisfied. The one or more update conditions may include one or more of: a particular time of day, a particular day of week, whether a threshold quantity of updates are available to update global ML model, or whether a threshold quantity of disparate updates are available to update global ML model.
In some implementations, the remote system may correspond to a remote server or a cluster of remote servers.
Various implementations can include a non-transitory computer readable storage medium storing instructions executable by one or more processors (e.g., central processing unit(s) (CPU(s)), graphics processing unit(s) (GPU(s)), digital signal processor(s) (DSP(s)), and/or tensor processing unit(s) (TPU(s)) to perform a method such as one or more of the methods described herein. Other implementations can include an automated assistant client device (e.g., a client device including at least an automated assistant interface for interfacing with cloud-based automated assistant component(s)) that includes processor(s) operable to execute stored instructions to perform a method, such as one or more of the methods described herein. Yet other implementations can include a system of one or more servers that include one or more processors operable to execute stored instructions to perform a method such as one or more of the methods described herein.
