Fast and slow synaptic plasticity enables concurrent control and learning
During many tasks the brain receives real-time feedback about performance. What should it do with that information, at the synaptic level, so that tasks can be performed as well as possible? The conventional answer is that it should learn by incrementally adjusting synaptic strengths. We show, however, that learning on its own is severely suboptimal. To maximize performance, synaptic plasticity should also operate on a much faster timescale - essentially, the synaptic weights should act as a control signal. We propose a normative plasticity rule that embodies this principle. In this, fast synaptic weight changes greedily suppress downstream errors, while slow synaptic weight changes implement statistically optimal learning. This enables near-perfect task performance immediately, efficient task execution on longer timescales, and confers robustness to noise and other perturbations. Applied in a cerebellar microcircuit model, the theory explains longstanding experimental observations and makes novel testable predictions.