In this innovative research, TopoLoss is introduced as a new loss function that promotes spatially organized topographic representations in AI models without compromising task performance. The TopoNets created using TopoLoss outperform previous models in terms of supervised topography, showcasing brain-like properties such as localized feature processing and increased efficiency. These models accurately predict brain responses and replicate key topographic signatures observed in the brain’s visual and language cortices. Overall, this work establishes a solid framework for integrating topography into top model architectures, pushing the boundaries of AI development to mirror the computational strategies of the human brain.
https://arxiv.org/abs/2501.16396