Where the brain first gains the ability to generalize—to extract relational structures that hold across changing contexts—remains an open question in AI and neuroscience. A new preprint, ReGraph, addresses this question with a computational account of the dual visual streams, the "what" and "where" pathways.
The authors argue that the foundation for this generalization capacity lies upstream of the hippocampus, rather than in the hippocampus itself. The model is designed to explain how context-invariant relational structures emerge within the visual system.
The abstract does not provide details on the model's architecture or experimental results, so the full account is not yet available from this source. The paper is posted on arXiv under identifier 2610.07962.