Toward a Simulated Avian Visual System: Connectome Mapping of the Pigeon Tectofugal Pathway as a Biologically-Grounded CNN Alternative

Abstract

Image recognition remains one of the central unsolved problems in machine learning. Despite superhuman benchmark accuracy, convolutional neural networks (CNNs) fail systematically at the tasks that define real-world deployment: they require orders of magnitude more labeled data than biological learners, collapse under adversarial perturbation, and degrade sharply under distribution shift. We argue that these failures are not engineering deficiencies but architectural ones — gradient descent over static feedforward graphs cannot recover the properties that evolution spent millions of years optimizing into biological visual systems. We propose a new class of visual inference engine: the Avian Visual Classification Circuit (AVCC), a computational simulation of the Columba livia (rock pigeon) visual-motor connectome, scoped to the tectofugal pathway and its beak-directed motor projections. The pigeon is selected not for convenience but for demonstrated performance: pigeons match or exceed state-of-the-art deep learning on categorical generalization, few-shot transfer, and medical image classification under matched data constraints — all without training. Grounding our approach in the recent Biological Processing Unit (BPU) proof-of-concept, we extend the paradigm to an organism purpose-built by evolution for the specific failure modes of CNNs. The AVCC accepts rasterized image inputs, processes them through biophysically realistic spiking neural network (SNN) dynamics derived from the reconstructed connectome, and emits classification outputs via a decoded pecking-motor readout — the only component trained via gradient descent. This proposal establishes priority of concept and defines a concrete research program toward a post-CNN visual inference architecture.