Toward biologically plausible predictive coding From spiking neurons to cortical microcircuits and cortico-subcortical loops

Predictive coding is a powerful computational framework proposing that the brain actively predicts sensory input and minimizes the resulting prediction errors. Despite its current spotlight in neuroscience—thanks to its ability to explain phenomena from neural responses to behavior—a concrete mapping onto real brain architecture and physiology remains a largely open question.
This thesis explores that question by examining predictive coding across multiple biological spatiotemporal scales. It begins by showing how predictive coding can be implemented with spiking neurons, preserving core inferential and generative capacities while introducing realistic neural dynamics and NMDA-receptor mediated synaptic plasticity. It then situates these computations within cortical microcircuits, identifying how excitatory and inhibitory cell types (pyramidal cells; PV, SST, VIP interneurons) and their laminar organization route prediction and error signals. Finally, the theory is extended beyond the cortex, integrating cortico-subcortical interactions to model how automatic responses interact with slower, context-sensitive control in perceptual decision-making.
Together, these studies demonstrate the viability of predictive coding through biologically constrained models that make concrete predictions about underlying neural mechanisms. While focused primarily on vision, the framework may extend naturally to other sensory modalities, offering a unifying perspective for how the brain integrates diverse sensory information into coherent perceptual experience. More broadly, predictive processing may provide a mechanistic rationale for the evolutionary emergence of a predictive brain—one that constructs hierarchical representations to interpret the world, guide adaptive behavior, and, perhaps, shed light on the neural foundations of consciousness.