A Model Cortical Network for Spatiotemporal Sequence Learning and Prediction

In this paper we developed a hierarchical network model, called Hierarchical Prediction Network (HPNet) to understand how spatiotemporal memories might be learned and encoded in a representational hierarchy for predicting future video frames. The model is inspired by the feedforward, feedback and lateral recurrent circuits in the mammalian hierarchical visual system. It assumes that spatiotemporal memories are encoded in the recurrent connections within each level and between different levels of the hierarchy. The model contains a feed-forward path that computes and encodes spatiotemporal features of successive complexity and a feedback path that projects interpretation from a higher level to the level below. Within each level, the feed-forward path and the feedback path intersect in a recurrent gated circuit that integrates their signals as well as the circuit's internal memory states to generate a prediction of the incoming signals. The network learns by comparing the incoming signals with its prediction, updating its internal model of the world by minimizing the prediction errors at each level of the hierarchy in the style of {\em predictive self-supervised learning}. The network processes data in blocks of video frames rather than a frame-to-frame basis. This allows it to learn relationships among movement patterns, yielding state-of-the-art performance in long range video sequence predictions in benchmark datasets. We observed that hierarchical interaction in the network introduces sensitivity to memories of global movement patterns even in the population representation of the units in the earliest level. Finally, we provided neurophysiological evidence, showing that neurons in the early visual cortex of awake monkeys exhibit very similar sensitivity and behaviors. These findings suggest that predictive self-supervised learning might be an important principle for representational learning in the visual cortex.