Predictive Sequence Learning in the Hippocampal Formation

Abstract

The interaction between hippocampus and cortex is key to memory formation and representation learning. Based on anatomical wiring and transmission delays, we proposed a self-supervised recurrent neural network (PredRAE) with a predictive reconstruction loss to account for the cognitive functions of hippocampus. This framework extends predictive coding in the time axis and incorporates predictive features in Bayes filters for temporal prediction. In simulations, we were able to reproduce characteristic place cell features such as one-shot plasticity, localized spatial representation and replay, which marks the trace of memory formation. The simulated place cells also exhibited precise spike timing, evidenced by phase precession. Trained on MNIST sequences, PredRAE learned the underlying temporal dependencies and a spontaneous representation of digit labels and rotation dynamics in the linear transformation of its hidden unit activities. Such learning is robust against 16 different image corruptions. Inspired by the brain circuit, the simple and concise framework has great potential to approximate human performance with its capacity to robustly disentangle representations and generalize temporal dynamics.

Sequential predictive learning is a unifying theory for hippocampal representation and replay

“…This past work has shown that: (1) Aspects of spatial tuning in the hippocampus are well-explained by predictive representations 11,55 , which can be learned from spatially tuned inputs with predictive Hebbian learning 79 , TD learning 80 , or spike-timing dependent plasticity 80,81 ; (2) Learning to predict memory embeddings from actions can link allocentric and egocentric representations 82 ; (3) Prediction can link spatially tuned inputs and relational structures 83 ; (4) Cloned hidden-Markov models that use discrete states to predict observations from actions can recapitulate many of the features of the hippocampus and support offline evaluations 84,85 ; (5) Engaging in path integration (i.e. predicting spatial locations from sequences of actions) leads to the emergence of grid cells and continuous attractor dynamics 48,50 ; (6) Training recurrent networks to predict hippocampal spiking data or linear place cell sequences can reproduce various features of hippocampal activity, including spike cross-correlations and sequence replay 86 . Our work demonstrates that these findings can be brought together under the umbrella of sequential predictive learning of sensory data in RNNs, and applied to ego-centric, high-dimensional, continuous sensory inputs.…”