Yi Ren

The paper identifies a scale-mixture Langevin sampling mechanism in recurrent cortical circuits, showing how nonlinear dynamics produce heavy-tailed statistics without explicit non-Gaussian components.

This study shows continuous attractor networks can simultaneously compute model evidence and posterior sampling in distinct subspaces, advancing understanding of neural computation.

The paper shows canonical cortical circuits naturally implement advanced Bayesian sampling algorithms, including natural gradient methods, through nonlinear dynamics and interneuron control.