Information retention by synaptic learning rules in stochastic spiking neural networks

Physical learning systems require a means to store information about the past. In the brain, information is thought to be stored in synaptic connections between neurons whose weights change in response to pre- and postsynaptic action potentials, or spikes, according to specific plasticity rules. To date, most models of synaptic plasticity are derived from in vitro studies in which spike times are under precise experimental control, but whether these models generalize to in vivo regimes of high spike time variability is not known. To shed light on this problem theoretically, here I estimate the mutual information between Poisson spike trains in a network of neurons and the synaptic weight updates they trigger via different plasticity rules. Although most synaptic plasticity models propose correlation- or coincidence-based rules based on the measurement of pre- and postsynaptic spike times, I show that such rules do not always maximize information storage. In the biologically important, energy-efficient operating regime of low firing rates and sparse plasticity, rules that simply count pre- and/or postsynaptic spikes in fact store more information per synapse than spike-timing-dependent rules that measure coincidence events. In such regimes, spikes are more common than coincidences hence better able to transmit information into synaptic weights. It may in turn be advantageous to encode information to be stored as spike counts, or time averages, rather than correlations. The theory suggests a normative account for recently discovered noncanonical hippocampal plasticity rules in vivo and predicts that spike-timing-dependent rules will be preferentially active in neural circuits operating at higher firing rates.