The thesis opens with a foundational chapter that distils the core principles of computational neuroscience —conductance-based neuron and synapse models, plasticity rules, and, crucially, a step-by-step derivation of gradient-based optimisation for large recurrent spiking neural networks (RSNNs). This tutorial lays the groundwork for understanding how biological constraints can coexist with modern training pipelines, serving as the methodological foundation for the visual-cortex studies that follow. Early vision is then recast as a dialogue between ascending sensory evidence and descending predictions. Re-examining the published connectome of the mouse primary visual cortex (V1) by Billeh et al. (2020), we find that its untrained wiring already embeds a signed mismatch code: brief halts in optic flow split layer-2/3 pyramidal neurons into
antagonistic depolarising and hyperpolarising ensembles, an emergent consequence of hard-wired asymmetries in thalamic drive and parvalbumin inhibition. Subsequently, we construct a biologically grounded V1 digital twin comprising∼200k point neurons, 201 cell types, and synaptic motifs derived from MICrONS electron microscopy, synaptic physiology, and Neuropixels recordings datasets. A multi-objective, exponentiated-gradient loss aligns firing-rate, orientation, and direction selectivity, as well as synchrony distributions with mouse recordings, while preserving the log-normal weight spectrum. The optimisation procedure spontaneously yields layer-specific, like-to-like excitation and deep anti-like inhibition, suggesting that canonical wiring rules can emerge from joint structural-functional pressure.
Finally, we close the loop by coupling the Billeh et al. V1 column to a size-matched lateromedial (LM) module through empirically derived, distance- and type-specific wiring rules. After targeted tuning, LM feedback (i) broadens V1 receptive fields, (ii) advances first-spike latencies by 2−4 ms, and (iii) boosts noisy-digit classification accuracy without inflating firing rates, indicating that feedback mainly acts as a gain-control/denoising signal rather than a feature generator.