Overfitting, Consciousness, and the Geometry of Experience: A Unified Computational Framework

Abstract

Abstract A central challenge in neuroscience and psychology is to explain how an individual’s unique life history sculpts their cognitive dispositions, giving rise to stable patterns of perception, thought, and action, as well as the phenomenal experience of consciousness. Here we propose a formal computational framework in which the unconscious mind is reconceptualised not as a passive repository of repressed content, but as an active, overfitted generative model of an individual’s experiential world. We show that the lifelong process of minimising variational free energy—formally equivalent to empirical risk minimisation in machine learning—inevitably leads to the over specialisation of internal model parameters θ to the particular statistical regularities of that individual’s sensory history. This overfitting creates a stable, low entropy informational manifold that determines the individual’s probabilistic “destiny”—their disposition to perceive and act. Within this framework, consciousness, including its social manifestations, is characterised as the recursive, energy consuming process of active inference, in which the brain dynamically minimises the prediction error between the overfitted prior and real time sensory data. By synthesising concepts from theoretical neuroscience, artificial intelligence, and non equilibrium thermodynamics, we derive a mathematical model of a “Consciousness Potential” and propose that the brain operates as an entropy reducing computational system governed by a fundamental information geometry. Our framework provides a unified mathematical language for describing the interplay among experience, disposition, and conscious awareness, and yields testable predictions for neuroimaging and artificial intelligence research.