How does the cortex of humans and primates interpret unstable and ambiguous sensory input in terms of lifelong prior experience? Current theories („free energy principle“, „predictive coding“) are based on equilibrium physics and posit reciprocal interactions between cortical levels („top-down“ and „bottom-up“) to compare sensory input with prior experience and minimise discrepancies.
I propose a novel hypothesis — an non-equilibrium hierarchy of birth-death processes — formulated as a mesoscopic model of cortical activity at level of columns. The model relies on standard cortical artchitecture and connectivity (feedforward projections shape tuning, recurring connections normalize gain) and is based on the dynamics of multi-stable perception, which strongly implicates a far-from-equilibrium birth-death process.
The model links all levels of analysis of Marr: computation (optimal inference), algorithm (birth-death hierarchy) and implementation (attractor dynamics of cortical columns). I conclude that that perceptual inference may rely on a variational principle of non-equilibrium („maximum caliber“) and that stochastic neural activity („noise correlations“ or „shared variability“) may be beneficial — not detrimental — for physiological function.
