Learning neural dynamics through instructive signals

Rapid learning is essential for robust behavior, allowing animals to quickly adapt to changing environments and task conditions. However, traditional Hebbian plasticity rules lack the power to induce the rapid, flexible changes in neural dynamics needed for such learning. Here we propose a network model with a rapid, heterosynaptic plasticity rule inspired by circuit motifs in hippocampus, cerebellum, and mushroom body, which relies on interactions between pre-synaptic activity and an "instructive signal" such as dopamine. We show that this rule can quickly produce highly flexible nonlinear dynamics using sparse, low-dimensional instructive signals while also admitting precise mathematical interpretation inspired by the well-known support vector machine. Using adaptive control theory, we additionally derive an online rule allowing such networks to learn using instructive signals guided by real-time feedback. Thus, heterosynaptic instruction-mediated plasticity is a powerful, biologically plausible and analytically tractable mechanism for the rapid learning of flexible dynamics, shedding light on the neural basis of complex adaptive behavior.