Predictive Sequence Learning in Recurrent Neocortical Circuits

Part of Advances in Neural Information Processing Systems 12 (NIPS 1999)

Bibtex Metadata Paper


Rajesh Rao, Terrence J. Sejnowski


Neocortical circuits are dominated by massive excitatory feedback: more than eighty percent of the synapses made by excitatory cortical neurons are onto other excitatory cortical neurons. Why is there such massive re(cid:173) current excitation in the neocortex and what is its role in cortical compu(cid:173) tation? Recent neurophysiological experiments have shown that the plas(cid:173) ticity of recurrent neocortical synapses is governed by a temporally asym(cid:173) metric Hebbian learning rule. We describe how such a rule may allow the cortex to modify recurrent synapses for prediction of input sequences. The goal is to predict the next cortical input from the recent past based on previous experience of similar input sequences. We show that a temporal difference learning rule for prediction used in conjunction with dendritic back-propagating action potentials reproduces the temporally asymmet(cid:173) ric Hebbian plasticity observed physiologically. Biophysical simulations demonstrate that a network of cortical neurons can learn to predict mov(cid:173) ing stimuli and develop direction selective responses as a consequence of learning. The space-time response properties of model neurons are shown to be similar to those of direction selective cells in alert monkey VI.