Abstract: An example assistant system can use a multimodal multitask medical machine-learned model to perform image processing to answer natural language queries. A device can process speech data or other natural language inputs to obtain a query. The query can be processed alongside image data that provides context for the query. The example system can receive a query associated with a particular task domain; generate, based on the query, a query input that comprises query instruction data from a first modality and query context data from a second modality; generate a combined input comprising the query input and an exemplar input, wherein the exemplar input comprises exemplar instruction data from the first modality and an exemplar context placeholder in lieu of exemplar context data from the second modality; process the combined input with a multimodal machine-learned model to generate output data; and output a query response based on the output data.

Abstract: An aspect of the present disclosure provides an example method comprising: receiving an input query associated with a particular task domain of a plurality of available task domains; obtaining a machine-learned prompt component and a curated prompt component, wherein the machine-learned prompt component comprises a plurality of machine-learned prompt values for the plurality of available task domains, and wherein the curated prompt component comprises a plurality of exemplar prompt values corresponding to one or more embedded natural language exemplars for the particular task domain from domain experts; and generating an output responsive to the input query by processing a combined prompt and the input query using a pre-trained machine-learned model, wherein the combined prompt comprises the machine-learned prompt component and the curated prompt component.

Abstract: Implementations disclosed herein are directed to various techniques for mitigating and/or preventing catastrophic forgetting in federated learning of global machine learning (ML) models. Implementations may identify a global ML model that is initially trained at a remote server based on a server data set, determine server-based data for global weight(s) of the global ML model, and transmit the global ML model and the server-based data to a plurality of client devices. The server-based data may include, for example, EWC loss term(s), client augmenting gradients, server augmenting gradients, and/or server-based data. Further, the plurality client devices may generate, based on processing corresponding predicted output and using the global ML model, and based on the server-based data, a corresponding client gradient, and transmit the corresponding client gradient to the remote server. Implementations may further generate an updated global ML model based on at least the corresponding client gradients.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a machine learning model having a set of local model parameters and a set of global model parameters under a partially local federated learning framework. One of the methods include maintaining local data and data defining the local model parameters; receiving data defining current values of the global model parameters; determining, based on the local data, the local model parameters, and the current values of the global model parameters, current values of the local model parameters; determining, based on the local data, the current values of the local model parameters, and the current values of the global model parameters, updated values of the global model parameters; generating, based on the updated values of the global model parameters, parameter update data defining an update to the global model parameters; and transmitting the parameter update data.