| May 2, 2024 | Our paper Stochastic Gradient Flow Dynamics of Test Risk and its Exact Solution for Weak Features, in collaboration with Anastasia Remizova and Nicolas Macris, has been accepted to ICML 2024. See you in Vienna in July! |
| Apr 15, 2024 | This week, I have the pleasure of attending and giving a talk at the From Theory to Practice workshop at the AIMS Research and Innovation Centre in Kigali, Rwanda. |
| Feb 12, 2024 | Understanding the role of stochasticity in gradient-bases learning dynamics, as well as in achieving different global minima and how it is related to good generalization performances, is of most importance. Based on a continuous-time representation of stochastic gradient descent, in our new preprint Stochastic Gradient Flow Dynamics of Test Risk and its Exact Solution for Weak Features with Anastasia Remizova and Nicolas Macris, we construct a path-integral framework to compute the difference between pure gradient flow and stochastic gradient flow trajectories in the limit of small learning rate. We successfully apply the formalism to a high-dimensional linear regression model with random projections. Check it out! |