Abstract: Application context may be interpolated between application state updates from structured and unstructured application state data. Irrelevant unimodal modules may be deactivated based on the structured application state data while relevant unimodal modules remain active. Unimodal features are generated from the unstructured application using the relevant modules. A neural module selection network module may be trained with a machine learning algorithm. Each unimodal modules may generate unimodal feature vectors from unstructured application data. A context state update module may determine which unimodal modules are irrelevant from structured application state data and deactivate the irrelevant modules but not the relevant ones. A multimodal neural network may take the active unimodal feature vectors and predict structured context data and send it to a uniform data system.

Abstract: A system for generating gameplay context information for a game may include a game screen classification module trained to classify contextually relevant data from gameplay data, one or more game object recognition modules trained to detect game icons from gameplay data, and a multimodal context generation neural network module trained to generate structured gameplay context information from the contextually relevant data and icons within the gameplay data. The multimodal context generation neural network module at least partially generates structured gameplay context information. The modules may include neural networks trained by suitable machine learning algorithms using suitable masked data and labeled data.

Abstract: Techniques for managing machine operations using encoded multi-scale time series data are provided. In one technique, operational data is received from a sensor coupled to an industrial device. For each portion in a first set of portions of the operational data (where each portion corresponds to a first time scale), first aggregated data is generated based on time series data from that portion and a first encoding is generated based on the first aggregated data. For each portion in a second set of portions of the operational data (where each portion of the second set corresponds to a second time scale that is different than the first time scale), second aggregated data is generated based on time series data from that portion and a second encoding is generated based on the second aggregated data. The operational data is classified to determine a condition of the industrial device during the time interval based on the first and second encodings.

Abstract: Techniques for managing machine operations using encoded multi-scale time series data are provided. In one technique, operational data is received from a sensor coupled to an industrial device. For each portion in a first set of portions of the operational data (where each portion corresponds to a first time scale), first aggregated data is generated based on time series data from that portion and a first encoding is generated based on the first aggregated data. For each portion in a second set of portions of the operational data (where each portion of the second set corresponds to a second time scale that is different than the first time scale), second aggregated data is generated based on time series data from that portion and a second encoding is generated based on the second aggregated data. The operational data is classified to determine a condition of the industrial device during the time interval based on the first and second encodings.