Shimizu & Toyoizumi's STP theory enables rapid temporal coding in neural networks

In a new paper on arXiv, researchers Genki Shimizu and Taro Toyoizumi from RIKEN Center for Brain Science present a normative information-theoretic theory of associative short-term synaptic plasticity (STP). Unlike traditional views that treat STP as a simple presynaptic filter independent of postsynaptic activity, recent experiments have shown that STP can be associative—depending on both pre- and postsynaptic coactivation. The authors extend Fisher-information-based learning to Tsodyks-Markram synapses, deriving learning rules for two key parameters: baseline synaptic weight and release probability. These rules operate under resource constraints to maximize the information encoded in neural firing about incoming stimuli.

The resulting model produces several striking predictions. For slowly varying inputs, the release-probability plasticity generates anti-causal connectivity—meaning synapses strengthen for inputs that precede spiking—and enhances response offset during sustained drive. After the stimulus is removed, the network exhibits reverse replay of activity sequences, a phenomenon observed in biological hippocampal place cells. Linear-response analysis reveals that STP introduces frequency-dependent phase selectivity, and constraints on release probability tune the temporal asymmetry of synaptic transmission. This positions release-probability plasticity as a principled, rapidly reconfigurable substrate for temporal coding, with potential implications for designing adaptive neural network architectures in AI and neuromorphic computing.