Abstract: Example implementations include methods, apparatuses, and computer-readable mediums of federated learning by a federated client device, comprising identifying client invariant information of a neural network for performing a machine learning (ML) task in a first domain known to a federated server. The implementations further comprising transmitting the client invariant information to the federated server, the federated server configured to generate a ML model for performing the ML task in a domain unknown to the federated server based on the client invariant information and other client invariant information of another neural network for performing the ML task in a second domain known to the federated server.

Abstract: A processor-implemented method for recognizing a natural language on a mobile device includes receiving an audio input. The method further includes using a neural network to generate local text corresponding to the audio input. The method still further includes generating a local confidence value for accuracy of the local text. The method includes transmitting, to a remote device, data corresponding to the audio input. The method further includes receiving remote text corresponding to the data, along with a remote confidence score for accuracy of the remote text. The method still further includes outputting the local text in response to the local confidence value being higher than the remote confidence score, and outputting the remote text in response to the remote confidence score being higher than the local confidence value.

Abstract: Embodiments include methods performed by a processor of a computing device for suggesting more efficient action sequences to a user. The methods may include recognizing a user action sequence including one or more user actions performed by the user to achieve a result, determining a first difficulty rating of the user action sequence, determining whether a cluster of multiple system action sequences exists within a cluster database in which each system action sequence of the one or more system action sequences produces the result. Methods may further include comparing the first difficulty rating to one or more difficulty ratings of the one or more system action sequences in response to determining that the cluster of multiple system action sequences exists within the cluster database, and displaying, via a display interface of the computing device, one or more system action sequences with a lower difficulty rating than the first difficulty rating.

Abstract: Systems and techniques are provided for training a neural network model or machine learning model. For example, a method of augmenting training data can include augmenting, based on a randomly initialized neural network, training data to generate augmented training data and aggregating data with a plurality of styles from the augmented training data to generate aggregated training data. The method can further include applying semantic-aware style fusion to the aggregated training data to generate fused training data and adding the fused training data as fictitious samples to the training data to generate updated training data for training the neural network model or machine learning model.

Abstract: Certain aspects of the present disclosure provide techniques for improved domain adaptation in machine learning. A feature tensor is generated by processing input data using a feature extractor. A first set of logits is generated by processing the feature tensor using a domain-agnostic classifier, and a second set of logits is generated by processing the feature tensor using a domain-specific classifier. A loss is computed based at least in part on the first set of logits and the second set of logits, where the loss includes a divergence loss component. The feature extractor, the domain-agnostic classifier, and the domain-specific classifier are refined using the loss.

Abstract: A processor-implemented method includes training a machine learning model on a source domain. The method also includes testing the machine learning model on a target domain, after training. The method further includes training the machine learning model on the target domain by regularizing weights of the machine learning model such that shift-agnostic weights are subjected to a higher penalty than shift-biased weights.

Abstract: In one embodiment, an electronic device includes an input device configured to provide an input stream, a first processing device, and a second processing device. The first processing device is configured to use a keyword-detection model to determine if the input stream comprises a keyword, wake up the second processing device in response to determining that a segment of the input stream comprises the keyword, and modify the keyword-detection model in response to a training input received from the second processing device. The second processing device is configured to use a first neural network to determine whether the segment of the input stream comprises the keyword and provide the training input to the first processing device in response to determining that the segment of the input stream does not comprise the keyword.

Abstract: Embodiment systems and methods for presenting a facial expression in a virtual meeting may include detecting a user facial expression of a user based on information received from a sensor of the computing device, determining whether the detected user facial expression is approved for presentation on an avatar in a virtual meeting, generating an avatar exhibiting a facial expression consistent with the detected user facial expression in response to determining that the detected user facial expression is approved for presentation on an avatar in the virtual meeting, generating an avatar exhibiting a facial expression that is approved for presentation in response to determining that the detected user facial expression is not approved for presentation on an avatar in the virtual meeting, and presenting the generated avatar in the virtual meeting.

Abstract: Embodiments include methods of assisting a user in locating a mobile device executed by a processor of the mobile device. Various embodiments may include a processor of a mobile device obtaining information useful for locating the mobile device from a sensor of the mobile device configured to obtain information regarding surroundings of the mobile device, anonymizing the obtained information to remove private information, and uploading the anonymized information to a remote server in response to determining that the mobile device may be misplaced. Anonymizing the obtained information may include removing speech from an audio input and compiling samples of ambient noise for inclusion in the anonymized information. Anonymizing the obtained information to remove private information includes editing an image captured by the mobile device to make images of detected individuals unrecognizable.

Abstract: Certain aspects of the present disclosure provide techniques for training a machine learning model. The method generally includes receiving, at a local device from a server, information defining a global version of a machine learning model. A local version of the machine learning model and a local center associated with the local version of the machine learning model are generated based on embeddings generated from local data at a client device and the global version of the machine learning model. A secure center different from the local center is generated based, at least in part, on information about secure centers shared by a plurality of other devices participating in a federated learning scheme. Information about the local version of the machine learning model and information about the secure center is transmitted by the local device to the server.

Abstract: Embodiment systems and methods for presenting a facial expression in a virtual meeting may include detecting a user facial expression of a user based on information received from a sensor of the computing device, determining whether the detected user facial expression is approved for presentation on an avatar in a virtual meeting, generating an avatar exhibiting a facial expression consistent with the detected user facial expression in response to determining that the detected user facial expression is approved for presentation on an avatar in the virtual meeting, generating an avatar exhibiting a facial expression that is approved for presentation in response to determining that the detected user facial expression is not approved for presentation on an avatar in the virtual meeting, and presenting the generated avatar in the virtual meeting.

Abstract: Certain aspects of the present disclosure provide a method of processing video data. In one example, the method includes receiving input video data; sampling a first subset of clips from the input video data; providing the first subset of clips to a first component of a machine learning model to generate first output; sampling a second subset of clips from the input video data, wherein the second subset of clips comprises fewer clips than the first subset of clips; providing the second subset of clips to a second component of the machine learning model to generate a second output; aggregating the first output from the first component of the machine learning model with the second output from the second component of the machine learning model to generate aggregated output; and determining a characteristic of the input video data based on the aggregated output.

Abstract: A device to process an audio signal representing input sound includes a hand detector configured to generate a first indication responsive to detection of at least a portion of a hand over at least a portion of the device. The device also includes an automatic speech recognition system configured to be activated, responsive to the first indication, to process the audio signal.

Abstract: Techniques for cross-domain adaptive learning are provided. A target domain feature extraction model is tuned from a source domain feature extraction model trained on a source data set, where the tuning is performed using a mask generation model trained on a target data set, and the tuning is performed using the target data set.

Abstract: A computer-implemented method of operating an artificial neural network for processing data having a frequency dimension includes receiving an input. The audio input may be separated into one or more subgroups along the frequency dimension. A normalization may be performed on each subgroup. The normalization for a first subgroup the normalization is performed independently of the normalization a second subgroups. An output such as a keyword detection indication, is generated based on the normalized subgroups.

Abstract: A method performed by an artificial neural network (ANN) includes determining, at a first stage of a multi-stage cross-attention model of the ANN, a first cross-correlation between a first representation of each modality of a number of modalities associated with a sequence of inputs. The method still further includes determining, at each second stage of one or more second stages of the multi-stage cross-attention model, a second cross-correlation between first attended representations of each modality. The method also includes generating a concatenated feature representation associated with a final second stage of the one or more second stages based on the second cross-correlation associated with the final second stage, the first attended representation of each modality, and the first representation of each modality. The method further includes determining a probability distribution between a set of background actions and a set of foreground actions from the concatenated feature representation.

Abstract: A device to process an audio signal representing input sound includes a user voice verifier configured to generate a first indication based on whether the audio signal represents a user's voice. The device includes a speaking target detector configured to generate a second indication based on whether the audio signal represents at least one of a command or a question. The device includes an activation signal unit configured to selectively generate an activation signal based on the first indication and the second indication. The device also includes an automatic speech recognition engine configured to be activated, responsive to the activation signal, to process the audio signal.

Abstract: A device to perform end-of-utterance detection includes a speaker vector extractor configured to receive a frame of an audio signal and to generate a speaker vector that corresponds to the frame. The device also includes an end-of-utterance detector configured to process the speaker vector and to generate an indicator that indicates whether the frame corresponds to an end of an utterance of a particular speaker.

Abstract: A device includes a screen and one or more processors configured to provide, at the screen, a graphical user interface (GUI) configured to display data associated with multiple devices on the screen. The GUI is also configured to illustrate a label and at least one control input for each device of the multiple devices. The GUI is also configured to provide feedback to a user. The feedback indicates that a verbal command is not recognized with an action to be performed. The GUI is also configured to provide instructions for the user on how to teach the one or more processors which action is to be performed in response to receiving the verbal command.

Abstract: In one embodiment, an electronic device includes an input device configured to provide an input stream, a first processing device, and a second processing device. The first processing device is configured to use a keyword-detection model to determine if the input stream comprises a keyword, wake up the second processing device in response to determining that a segment of the input stream comprises the keyword, and modify the keyword-detection model in response to a training input received from the second processing device. The second processing device is configured to use a first neural network to determine whether the segment of the input stream comprises the keyword and provide the training input to the first processing device in response to determining that the segment of the input stream does not comprise the keyword.