#20 · Systems biology

Normative and random network models account for sensory and neural population data

The dominant thread is formal, model-based explanation of sensory and neural computation: each item derives or extends a framework and assesses it against data or structural constraints. A variational free-energy extension of predictive coding under exponential-family assumptions yields nonlinear dynamics with local plasticity. A recurrent hierarchical temporal-prediction model reproduces dorsal visual pathway tuning and best fits neural responses among the normative models considered. A relay model extends the race architecture with cross-modal initiation, reproducing redundant-signal speed-ups and race-inequality violations in multisensory response times. Two preprints address connectivity and population structure: a maximum-entropy framework derives connectivity from task constraints matching gradient-descent-trained networks, and random network models match observed population dimensionality. What is new is temporal prediction and cross-modal initiation incorporated into established frameworks.