Discovering plasticity rules that organize and maintain neural circuits

Biological neural networks exhibit remarkable computational abilities despite their heterogeneous and evolving synaptic structures. While prior research has explored the role of plasticity rules in organizing network function, it is uncertain whether these rules remain effective amidst structural fluctuations. Drawing inspiration from a recent study in zebra finch demonstrating the resilience of the song premotor area HVC after significant cell perturbation (Wang et al, 2022), we employ a meta-learning approach to find robust sets of unsupervised, online rules that can organize and maintain the dynamics of a similar circuit. Our meta-learned rules permit the network to accurately represent the time since it received an initial stimulus by organizing a sequential structure and significantly outperform another model of sequence formation when synapses are continually turned over. Further, as HVC possesses a large population of interneurons, we ask what the utility of plasticity on excitatory (E) to inhibitory (I) and I to E synapses might be as the network is continually altered and find that I to E plasticity regulating the firing rate of the postsynaptic cell can improve encoding of time by allowing the network to compensate for inhomogeneity in its feedforward structure.