Abstract
Information processing in the brain relies on the transmission of spikes through chemical synapses whose efficacies often depend on their recent firing history. While effects of such short-term plasticity on neural information processing have long been studied, it is unclear how interactions between short-term and long-term plasticity affect the learning abilities of neural networks. Here, we show that long-term changes to the short-term plasticity of individual synaptic connections enable neurons to learn to process temporal sequences of spikes as if they were spatial patterns. This mechanism allows neural circuits to flexibly increase their capacity and robustness at the expense of elevated spiking activity. We further show that neurons with plastic short-term plasticity can learn to discriminate between inputs on the basis of multineuronal spike correlations that extend over space and time. Our model fits recent electrophysiological measurements of short-term plasticity in the mouse and human neocortex and is consistent with the distributions of short-term plasticity induction and recovery. Our theory predicts that the learning rule observed at a given synaptic connection depends on the degree and type of short-term plasticity which is induced by the induction protocol for long-term plasticity.