TopoAudio's topographic training creates brain-like auditory AI components

A new preprint from Haider Al-Tahan and colleagues introduces TopoAudio, a family of topographic auditory models that explicitly replicate the spatial organizing principle of brain cortices. The models add two constraints during training: a wiring-length penalty that keeps connections short on a simulated two-dimensional cortical sheet, and a smoothness objective that encourages nearby units to develop similar response tunings. These constraints are grounded in the known anatomy of the human auditory cortex, where neurons with similar sound preferences cluster together. The paper (arXiv:2509.24039) argues that if topography is a fundamental feature of the brain, it should shape not just physical layout but the internal structure of neural population codes.

Despite the heavy spatial constraints, TopoAudio matches non-topographic baselines on standard tasks: speech recognition and environmental sound classification, as well as predicting human fMRI responses to natural sounds. The key difference emerges at the component level. Using dimensionality-reduction techniques to decompose model representations into interpretable sound-category components (speech, music, song), the team found TopoAudio produces more compact, sharper components that align closely with the same decompositions derived from human ECoG recordings. Standard models also show components, but they are messier and less consistent with brain data. This points to topography as a general inductive bias, not a performance tradeoff. The authors propose component-level alignment as a complementary metric for evaluating model-brain correspondence, particularly for auditory processing, and release code for their project page.