Modelling Predictive Coding in the Primary Visual Cortex (V1): Layer 4 Receptive Field Properties in a Balanced Recurrent Spiking Neuronal Network

Abstract

Understanding how the cortex encodes sensory input in a biologically efficient and computationally robust manner remains a central question in neuroscience. Predictive coding offers a compelling theoretical framework for such cortical processing, but existing models lack the biological detail to fully explain the function of the cortical microcircuits. This study introduces a spiking neural network model of layer 4 of the primary visual cortex (V1), grounded in predictive coding principles, to clarify how the thalamorecipient layer transforms feedforward input into prediction-error-like signals under realistic excitatory-inhibitory constraints and to yield testable circuit-level predictions. The model integrates structured feedforward input, distinct excitatory and inhibitory populations, and balanced lateral connectivity to simulate spontaneous and stimulus-driven activity. Network responses are systematically examined under spatially unstructured noise input and structured grating stimuli. Neural membrane potentials encode real-time reconstruction errors between external input and internal estimates, with spikes dynamically correcting these mismatches. The network reproduces hallmark in vivo features, including irregular spontaneous activity, sparse and selective responses, and emergent orientation and phase tuning. Excitatory-Inhibitory (E-I) balance was maintained across conditions, with inhibitory neurons exhibiting tighter input coupling than excitatory neurons. Furthermore, the network exhibited contrast-dependent modulation of firing rates and E-I balance, dynamically adjusting its activity to changes in input strength. Decoding analyses demonstrates that structured inputs can be robustly reconstructed under moderate noise levels, although decoding fidelity declines sharply under severe corruption. Together, these results suggest that cortical layer 4 may serve as a structured sensory encoding stage in a hierarchical predictive coding system, providing a biologically grounded foundation for modeling prediction error computations in higher cortical areas. Author summaryIn this Study, we present a biologically grounded spiking neural network model of layer 4 of the primary visual cortex, built within the predictive coding framework. The aim is to better understand how this early cortical layer encodes sensory information while maintaining realistic neural dynamics. Many predictive coding models focus on higher cortical layers 2/3 and overlook layer 4s role. To address this, we develop here a network that integrates structured feedforward input via Gaborfiltered receptive fields, distinct excitatory and inhibitory populations, and fixed lateral connectivity, all adhering to Dales law. The model reproduces several in vivo features observed in layer 4 of the visual cortex, including sparse, irregular spiking, emergent orientation and phase tuning, and contrast-dependent firing. Notably, excitation and inhibition are dynamically balanced across input conditions without requiring synaptic learning. We also show that decoding performance remains robust under moderate noise levels, supporting that layer 4 provides a stable sensory foundation for higher-level prediction. This model offers a biologically realistic implementation of prediction error computation and sets the stage for hierarchical extensions that include feedback and learning. Overall, this work provides insights into how structured sensory representations and balance emerge in cortical microcircuits through architecture alone.