The Geometry of Cortical Computation: Manifold Disentanglement and Predictive Dynamics in VCNet

Brennen Hill, Zhang Xinyu, Timothy Putra Prasetio

Keywords: Representational Geometry, Neural Manifolds, Manifold Disentanglement, Predictive Coding, Dynamical Systems, Neuro-inspired Architecture, Equivariant Representations, Hierarchical Representations, Geometric Deep Learning, Computational Neuroscience.

TL;DR: VCNet models the primate visual cortex as a geometric dynamical system, framing its dual-stream processing and predictive feedback as manifold disentanglement and geodesic refinement to learn robust, structured representations.

Abstract: Modern convolutional neural networks (CNNs) face fundamental limitations such as data inefficiency, poor out-of-distribution generalization, and vulnerability to adversarial perturbations. These shortcomings stem from a lack of inductive biases reflecting the visual world's inherent geometric structure. In contrast, the primate visual system offers a blueprint for more capable artificial vision; its efficiency and robustness derive from architectural principles evolved to internalize these structures. This paper introduces the Visual Cortex Network (VCNet), a novel architecture inspired by the macro-scale organization of the primate visual cortex. VCNet emulates key biological mechanisms within a geometric framework: hierarchical processing, dual-stream segregation for learning disentangled representations, and top-down predictive feedback for refinement. Interpreted through geometry and dynamical systems, we posit these mechanisms guide the learning of structured, low-dimensional neural manifolds. We evaluate VCNet on two specialized benchmarks: the Spots-10 animal pattern dataset to probe natural texture sensitivity, and the Stanford Light Field dataset to test performance on higher-dimensional visual data. VCNet surpasses contemporary models of comparable size, achieving 92.1% accuracy on Spots-10 and 74.4% on the light field dataset. This work demonstrates that integrating neuroscientific principles through a geometric lens can lead to more efficient, robust models, offering a promising direction for machine learning.

Submission Number: 7