UNSUPERVISED LEARNING TECHNIQUES FOR TEMPORAL DIFFERENCE MODELS

摘要

A temporal difference model can be trained to receive at least a first state representation and a second state representation that respectively describe a state of an object at two different times and, in response, output a temporal difference representation that encodes changes in the object between the two different times. To train the model, the temporal difference model can be combined with a prediction model that, given the temporal difference representation and the first state representation, seeks to predict or otherwise reconstruct the second state representation. The temporal difference model can be trained on a loss value that represents a difference between the second state representation and the prediction of the second state representation. In such fashion, unlabeled data can be used to train the temporal difference model to provide a temporal difference representation. The present disclosure further provides example uses for such temporal difference models once trained.