Neural Learning of Temporal Structures

Abstract

This project concerns the importance of learning fine temporal structures of uncertain events on judgment and decision-making. We claim that even for a completely random binary sequence (i.e., sequence of Bernoulli trials of p=0.5), there are richer temporal structures embedded in it, and that the humans seem to be able to pick them up and manifest them in observed behavior. This presents a challenge to current machine learning systems. Even with a sophisticated deep learning network well-tailored for temporal sequence learning (e.g., RNN and LSTM), learning to predict completely random sequence can be shown to be futile - there is simply nothing naively probabilistic there to learn from in the first place.