Xiaoliang Luo

The paper shows that brain regions and deep learning models stretch representations along task-relevant dimensions to optimize performance through error minimization.

The paper argues that spatial cell types in the brain may not be essential for spatial cognition, as similar patterns emerge in non-spatial deep neural networks.

The paper proposes a controller-peripheral architecture to model brain coordination for learning, showing how higher-level control shapes lower-level representations and outperforms gradient descent models.