Abstract: A method includes obtaining first feature vectors and second feature vectors representing contents of a first and second image frame, respectively, of an input video. The method may also include generating, based on the first feature vectors, first slot vectors, where each slot vector represents attributes of a corresponding entity as represented in the first image frame, and generating, based on the first slot vectors, predicted slot vectors including a corresponding predicted slot vector that represents a transition of the attributes of the corresponding entity from the first to the second image frame. The method may additionally include generating, based on the predicted slot vectors and the second feature vectors, second slot vectors including a corresponding slot vector that represents the attributes of the corresponding entity as represented in the second image frame, and determining an output based on the predicted slot vectors or the second slot vectors.

Abstract: Generally, the present disclosure is directed to novel machine-learned classification models that operate with hard attention to make discrete attention actions. The present disclosure also provides a self-supervised pre-training procedure that initializes the model to a state with more frequent rewards. Given only the ground truth classification labels for a set of training inputs (e.g., images), the proposed models are able to learn a policy over discrete attention locations that identifies certain portions of the input (e.g., patches of the images) that are relevant to the classification. In such fashion, the models are able to provide high accuracy classifications while also providing an explicit and interpretable basis for the decision.

Abstract: The present disclosure provides a neural network including one or more layers with relaxed spatial invariance. Each of the one or more layers can be configured to receive a respective layer input. Each of the one or more layers can be configured to convolve a plurality of different kernels against the respective layer input to generate a plurality of intermediate outputs, each of the plurality of intermediate outputs having a plurality of portions. Each of the one or more layers can be configured to apply, for each of the plurality of intermediate outputs, a respective plurality of weights respectively associated with the plurality of portions to generate a respective weighted output. Each of the one or more layers can be configured to generate a respective layer output based on the weighted outputs.

Abstract: Generally, the present disclosure is directed to novel machine-learned classification models that operate with hard attention to make discrete attention actions. The present disclosure also provides a self-supervised pre-training procedure that initializes the model to a state with more frequent rewards. Given only the ground truth classification labels for a set of training inputs (e.g., images), the proposed models are able to learn a policy over discrete attention locations that identifies certain portions of the input (e.g., patches of the images) that are relevant to the classification. In such fashion, the models are able to provide high accuracy classifications while also providing an explicit and interpretable basis for the decision.