Non-Hebbian properties of long-term potentiation enable high-capacity encoding of temporal sequences.

Abstract

A hypothesis commonly found in biological and computational studies of synaptic plasticity embodies a version of the 1949 postulate of Hebb that coactivity of pre- and postsynaptic elements results in increased efficacy of their synaptic contacts. This general proposal presaged the identification of the first and still only known long-lasting synaptic plasticity mechanism, long-term potentiation (LTP). Yet the detailed physiology of LTP induction and expression differs in many specifics from Hebb's rule. Incorporation of these physiological LTP constraints into a simple non-Hebbian network model enabled development of "sequence detectors" that respond preferentially to the sequences on which they were trained. The network was found to have unexpected capacity (e.g., 50 x 10(6) random sequences in a network of 10(5) cells), which scales linearly with network size, thereby addressing the question of memory capacity in brain circuitry of realistic size.