From Recurrent Adaptation to Hebbian Plasticity: Biologically Plausible Networks of Typical and Atypical Sensory Processing

Abstract

Most computational approaches utilized to study the brain today avoid explicit incorporation of biologically plausible mechanisms, instead prioritizing performance benchmarks. While these networks might display responses like those exhibited by animals, it is often by converging on mechanistic solutions that lack biological correlates. In doing so, findings derived from such models cannot be directly validated against physiological recordings, and inferences drawn from such a network might thereby not hold true mechanistic value. Recognizing this hurdle, a growing number of recent studies have adopted biophysical mechanisms into their computational simulations, resulting in findings that account for perceptual phenomena that cannot be studied by traditional computational approaches alone. Following a similar strategy, this thesis explores two studies designed with enough resolution to reproduce specific biological phenomena while at the same time remaining computationally tractable. The first study introduces AdaptNet, a motion processing network that learns from natural sequences while implementing neuronal adaptation — a mechanism long implicated in efficient coding and perceptual aftereffects. The second project builds a spiking model of the superior colliculus (SC) with explicit AMPA/NMDA conductances, GABA inhibition, and spike timing dependent plasticity. The study initially validates the network’s responses against established metrics of multisensory integration and then analyses how perturbations in the model’s mechanics lead to the altered responses observed in conditions like autism spectrum disorder (ASD). Taken together, these models advocate for ‘minimal realism’ — careful adherence to key biologically grounded mechanisms, when balanced alongside planned abstraction of secondary mechanisms, can produce network architectures that can be used to derive useful insights about neural activity as well as behavioural responses, in both health and dysfunction.