Structured Sparsification of Signal-Transmission Networks Enhances Visual Information Coding

AO_SCPLOWBSTRACTC_SCPLOWSensory coding depends on the architecture of cortical neural networks, yet the principles linking network organization to coding performance remain poorly understood. To address this question, we inferred spiking signal-transmission networks at single-neuron resolution from simultaneous Neuropixels recordings across six visual areas and related their organization to the fidelity and speed of sensory coding. Improved coding was associated with structured sparsification of these networks, characterized by fewer, more local, modular, feature-specific, and feedforward connections. To identify the mechanisms through which structured sparsification enhances coding, we used rate-based network models within a linear Fisher information framework. Our results show that structured sparsification improves coding fidelity by reducing shared variability, decreasing signal-noise alignment, and sharpening neuronal selectivity. It also accelerates coding by promoting more feedforward transmission of sensory signals. Together, these findings provide a mechanistic link between the architecture of cortical signal-transmission networks and the fidelity and speed of population codes.