TopoNets: High-Performing Vision and Language Models Mimicking Brain Topography
2025-02-03
Researchers introduce TopoLoss, a novel method for incorporating brain-like topography into leading AI architectures (convolutional networks and transformers) with minimal performance loss. The resulting TopoNets achieve state-of-the-art performance among supervised topographic neural networks. TopoLoss is easy to implement, and experiments show TopoNets maintain high performance while exhibiting brain-like spatial organization. Furthermore, TopoNets yield sparse, parameter-efficient language models and demonstrate brain-mimicking region selectivity in image recognition and temporal integration windows in language models, mirroring patterns observed in the visual cortex and language processing areas of the brain.
Read more
AI