A Renormalization-Based Approximation Theory for a Unified Computational Model of the Human Temporal Cortex: From Multimodal Abstraction to Scaled-Down Present-Day Systems

This work develops a mathematically grounded approximation theory for a unified computational model of the human temporal cortex. Building on a previously proposed multimodal architecture spanning STG, STS, MTG, and ATL, the paper introduces a renormalization-based framework that systematically derives reduced models at multiple scales. The theory formalizes how rate–distortion optimization, renormalization-group (RG) dynamics, and group-theoretic invariance jointly constrain representational structure across the temporal hierarchy. The resulting approximation sequence provides a principled account of how unimodal sensory encoding, cross-modal integration, and multimodal conceptual abstraction emerge as scale-dependent regimes of a single underlying model. The framework also reveals a bidirectional developmental logic: the full architecture can be approached bottom-up (from sensory systems toward abstraction) or top-down (from abstract multimodal representations toward perceptual grounding). This dual perspective clarifies the relationship between biological computation and modern multimodal AI systems, which can be interpreted as coarse-grained approximations of the full temporal-cortex model. The paper includes mathematical definitions of approximation errors, a unified approximation criterion, worked examples, and appendices providing formal justification of the framework. It establishes the theoretical foundations required for future empirical, computational, and applied developments.