Optimisation of learning-to-learn in spiking neural circuits

Abstract

Situated at the intersection of artificial intelligence and theoretical neuroscience, spiking neural networks (SNNs) have proven valuable for modelling and predicting neural phenomena. SNNs provide rich dynamics that are not replicable by conventional neural networks, exhibiting desirable properties such as self-healing, the ability to exploit noise as a resource, and the ability to solve difficult constraint problems like Sudoku or NP-hard problems such as the Travelling Salesman Problem without training. This work aims to discover initial configurations that lead to more effective and generalisable learning as part of a broader effort to deduce the computational principles that achieve generalisable learning in the human brain. Building on Pecevski & Maass, the work explores density estimation in simulations of SNN winner-take-all (WTA) circuits and similar constructions by using genetic algorithms (GA) and natural evolution strategies (NES) to search for individual optimal configurations that minimise Kullback-Leibler (KL) divergence between multiple estimate and target distributions. It demonstrates that in some tasks optimal network configurations outperform published results even though theoretical zero-delay conditions do not hold, that generalisability can be achieved with a significant but non-fatal impact in both conditional and joint estimation tasks, that biological plausibility can be pushed by introducing a novel architecture with realistic modifications that also achieves competitive performance in conditional estimation tasks, and that there exists a negative relationship between synaptic delay and estimation performance.