Structured connectivity for structured sequential computations

Abstract

Cortical computations emerge from the coordinated activity of recurrent neural circuits and are often described through population dynamics and neuronal selectivity. However, how the backbone of the computations - the circuit connectivity - gives rise to these dynamics and organizes the computations, remains largely unknown. Mixed selectivity, which can arise without structured connectivity, has been proposed as a substrate for flexible computation, but alone it does not account particularly for sequential computations, where distinct selectivities have been observed. Here, we investigate this problem using the delayed match-to-category (DMC) task, which requires categorization, memory maintenance, comparison, and decision-making, and in which pure, mixed, and time-varying selectivity have been observed. Using explicit connectivity models, we identify circuit motifs that generate these selectivity types and interconnect them to implement the computations required for behavior. Within these motifs, mixed selectivity neurons implement comparison via localized, distributed, or hybrid representations, whereas time-varying selectivity neurons gate the activation and timing of output. Recurrent networks trained from random initial connectivity develop similar motifs, demonstrating that learning leads to structured connectivity to support computations. The framework generalizes to extended task variants, revealing reusable circuit principles that mechanistically link connectivity, neuronal selectivity, and sequential computation.