AI model spontaneously recreates primate visual cortex direction maps
Self-supervised learning yields brain-like MT topography without explicit objective
A team led by Zhaotian Gu trained a spatiotemporal TDANN (Topographic Deep Artificial Neural Network) using a 3D ResNet backbone on naturalistic videos with Momentum Contrast (MoCo) self-supervised learning. To enforce topographic organization, they added a biologically inspired spatial loss that encourages neighboring neurons to have similar receptive fields. The result: the model autonomously formed direction-selective maps and topological pinwheel structures that quantitatively match in vivo recordings from macaque middle temporal (MT) cortex. Key metrics—direction selectivity index, circular variance, and pinwheel density—all fell within physiological baselines.
Crucially, the paper reveals that MT tuning properties arise from a strict optimization trade-off between discriminative pressure (task performance on video understanding) and spatial regularization (topographic smoothness). This is the first time a computational model has explained both ventral stream (object recognition) and dorsal stream (motion processing) topographies under a single principle. The work bridges neuroscience and AI, suggesting that self-supervised spatiotemporal learning may be the underlying driver of cortical map formation in the brain.
- Model: spatiotemporal TDANN with 3D ResNet trained via MoCo self-supervision on natural videos
- Emergent MT direction maps match macaque physiology: selectivity index, circular variance, pinwheel density
- Reveals a trade-off between task-driven learning and spatial regularization as the origin of MT topography
Why It Matters
Unifies ventral and dorsal stream computational theories, enabling brain-like AI vision and insights into cortical self-organization.