Discovering plasticity rules for learning and resilience in neural circuits

Abstract

While modern supervised and reinforcement learning algorithms can train neural networksto solve a wide range of tasks, the brain often operates in data-sparse regimes where such extensive supervision is unavailable. This thesis argues that the brain succeeds in these settings by leveraging inductive biases about the tasks it is likely to encounter. These biases are embedded in initial connectivity, cell-type structure, and critically, in synaptic plasticity rules. Here, we investigate how unsupervised synaptic plasticity can shape neural circuits prior to extensive behavioral experience. In the first part of this thesis, we study plasticity in the zebra finch song system. In collaboration with researchers at the California Institute of Technology, we examine the restoration of singing behavior following viral perturbation of nucleus HVC, a premotor region essential for song production. Adult male zebra finches transiently lose song after viral manipulation but recover within approximately two weeks. Strikingly, birds prevented from practicing during early recovery subsequently require less practice to regain song, suggesting that recovery is partially unsupervised. We model this process using several unsupervised plasticity mechanisms, including spike timing-dependent and homeostatic plasticity. While standard homeostatic rules restore regular spiking activity in a network model of HVC, they fail to reproduce experimentally observed synaptic reorganization. We therefore propose a local population-level homeostatic rule that recruits previously silent neurons, accounting for both activity recovery and synaptic changes. In the second chapter, we employ meta-learning, a technique by which biologically plausible learning rules are learned via a supervised procedure, to discover biologically plausible plasticity rules that organize robust sequential dynamics in HVC-like networks. In this framework, candidate unsupervised plasticity rules are optimized by a supervised outer loop to maximize a task objective. Starting from disordered connectivity, the learned rules reliably self-organize networks into sequence-generating circuits resembling those observed in vivo. Analysis of resulting rules reveals that plasticity on recurrent excitatory synapses generalizes Oja’s rule, replacing the classical Hebbian term with a spike timing-dependent component. We further show that learned plasticity rules can compensate for continual synaptic turnover and that learned inhibitory plasticity enhances the precision and robustness of sequential dynamics. In the final chapter, we apply meta-learning to the self-organization of neural inte- grators—circuits that generate long timescales via carefully tuned structure to maintain representations of sensory inputs. Such integrators underlie functions including head di- rection coding and oculomotor control. We hypothesize that unsupervised plasticity can shape these circuits from weak structural priors. Using meta-learning, we identify plasticity rules that reliably organize integration dynamics without requiring previously hypothesized anti-Hebbian mechanisms. Instead, the learned rules rely heavily on three-factor plasticity. In a simplified model, we demonstrate how such three-factor mechanisms can tune integrator circuitry and stabilize persistent dynamics.