Self-organized MT Direction Maps Emerge from Spatiotemporal Contrastive Optimization
Researchers have developed a spatiotemporal Topographic Deep Artificial Neural Network (TDANN) to understand the organization of the primate visual cortex's dorsal stream. By training a 3D ResNet on natural videos using a self-supervised contrastive learning method and a spatial loss, the model spontaneously generated direction maps and topological structures similar to those found in the middle temporal (MT) area. The study suggests that the specific tuning properties of MT neurons arise from a balance between discriminative pressures and spatial regularization, unifying computational principles across different visual processing streams. AI
IMPACT This research offers a computational framework that could advance understanding of brain organization and potentially inform future AI architectures for visual processing.