Longsheng Jiang

This paper investigates working memory as a form of recurrent fast-weight computation, showing that a model based on linear Transformer computation can replicate neural dynamics observed in macaque prefrontal cortex during a visuospatial task.

The paper shows that stimulus-driven changes in neuronal connectivity and thermodynamic states lead to more efficient coding in mouse V1 by reorganizing neural activity onto low-dimensional manifolds.

The paper investigates how the brain categorizes non-linearly separable stimuli by analyzing neural manifolds in macaques during a motion-induced illusory contour classification task, revealing a transformation from sensory to perceptual manifolds through geometric operations.

The paper analyzes single-neuron network topology in mouse V1 and finds it exhibits extreme small-world, modular, and rich-club properties compared to macroscopic brain networks.

The paper shows that neural geometries in the dorsal prefrontal cortex transform from an asymmetric sensory manifold to a symmetric intermediate state and then to an asymmetric mnemonic manifold during a visuospatial working memory task.