Evolutionary Optimization of Biologically-Inspired Mechanistic Input-Output Neural Circuit Models in the Auditory Pathway

Understanding how to compute specific functions from neural architecture is a fundamental challenge in neuroscience, with profound implications for artificial intelligence, neuromorphic computing, and neurological health. This study employs evolutionary meta-heuristic algorithms to optimize a mechanistic input-output spiking neural network (SNN) model of the cochlear nucleus, a key circuit in the auditory pathway. The research integrates biologically inspired initialization, iterative optimization, and data-driven validation to explore how connectivity and topology influence input-output relationships in neural circuits. Synthetic input-output spike train pairs for spiral ganglion and bushy cells are generated using a computational model of the cochlear nucleus. SNNs are initialized with biologically plausible parameters, and their synaptic weights and delays are refined using spike-timing dependent plasticity (STDP) while connectivity and distribution of cellular properties are evolved over generations. Fitness is evaluated based on the accuracy of replicating experimentally observed input-output patterns. The optimization is performed in parallel across a population of networks to enhance computational efficiency, with characteristic motifs in the optimized architectures analyzed to uncover their functional significance. This approach not only seeks to replicate the computations of biological circuits with high fidelity but also provides insights into the principles underlying neural circuit architecture. Preliminary findings are expected to demonstrate that SNNs can accurately mimic biological computations, offering a scalable methodology for computational neuroscience. SNNs are particularly promising for designing energy efficient models which are capable of reducing the carbon footprint of machine learning, by mimicking the sparse, event driven nature of biological neural activity. This work also has the potential to reduce the computational and time demands of traditional connectomic reconstruction by learning circuit architectures directly from empirical input-output data. Additionally, these results could inform the design of energy efficient neuromorphic hardware, test hypotheses about auditory processing, and guide neural architecture search (NAS) methods for artificial neural networks.