Speakers > Sophie Denève
Sophie Deneve is a CNRS Director of research and a computational neuroscientist. After a PhD in Rochester and a Research fellowship in Gatbsy Computational Neuroscience unit, London, she joined the Cognitive and Computational Neuroscience laboratory in Ecole Normale Supérieure, Paris. Her research centers on how neural structures can learn and implement predictive (i.e. generative) models of the world, and what crucial role spiking dynamics and excitatory to inhibitory balance (ubiquitous features of biological networks) could play in these processes.
Learning generative models of latent dynamics in spiking neural networks
Our neural circuits, evolved under biophysical and energy constraints, support learning time-varying sensorimotor and cognitive computations exhibiting remarkable generalization capabilities and flexibility. It is unclear how these circuits, given few example trajectories, can stably learn a robust dynamical model that generalizes well. Here, we introduce a biologically plausible learning framework that enables model cortical neurons to stably and robustly solve this overfitting problem. In this framework, spiking and learning dynamics decrease a Lyapunov function guiding stable learning while trading-off coding accuracy for cost, resulting in a spike-by-spike representational regularization. The framework outperforms alternative spiking representations, solves generalization and classification of human motion dynamics, and learns to perform a flexible cognitive task by decomposing it into low dimensional subspaces of the high-dimensional neural state space. |
