IC Colloquium: Optimization and Monte Carlo Sampling in Large Scale Machine Learning
Event details
| Date | 14.03.2019 |
| Hour | 10:15 › 11:15 |
| Location | |
| Category | Conferences - Seminars |
By: Nicolas Flammarion - UC Berkeley
IC Faculty candidate
Abstract:
Optimization algorithms and Monte Carlo sampling algorithms have provided the computational foundations for the rapid growth in applications of machine learning in recent years. In this talk we will see how the interplay between gradients, stochastic and acceleration can be exploited to obtain fast and efficient statistical procedures.
To begin we present the first algorithm which achieves jointly optimal prediction rates both in term of computational speed and statistical efficiency. This new algorithm is based on combining both averaging and acceleration.
Second, we show how sampling algorithms can be seen as optimization algorithms in the space of measures. Using this perspective, we interpret the underdamped Langevin algorithm as performing accelerated gradient descent on the KL divergence. The power of this approach allows us to obtain faster convergence rates for non convex sampling problems.
Bio:
Nicolas Flammarion is a postdoctoral fellow in computer science at UC Berkeley, hosted by Michael I. Jordan. He received his PhD in 2017 from Ecole Normale Supérieur in Paris, where he was advised by Alexandre d'Aspremont and Francis Bach. In 2018, he received the prize of the Fondation Mathematique Jacques Hadamard for the best PhD thesis in the field of optimization. His research focuses primarily on learning problems at the interface of machine learning and optimization.
More information
IC Faculty candidate
Abstract:
Optimization algorithms and Monte Carlo sampling algorithms have provided the computational foundations for the rapid growth in applications of machine learning in recent years. In this talk we will see how the interplay between gradients, stochastic and acceleration can be exploited to obtain fast and efficient statistical procedures.
To begin we present the first algorithm which achieves jointly optimal prediction rates both in term of computational speed and statistical efficiency. This new algorithm is based on combining both averaging and acceleration.
Second, we show how sampling algorithms can be seen as optimization algorithms in the space of measures. Using this perspective, we interpret the underdamped Langevin algorithm as performing accelerated gradient descent on the KL divergence. The power of this approach allows us to obtain faster convergence rates for non convex sampling problems.
Bio:
Nicolas Flammarion is a postdoctoral fellow in computer science at UC Berkeley, hosted by Michael I. Jordan. He received his PhD in 2017 from Ecole Normale Supérieur in Paris, where he was advised by Alexandre d'Aspremont and Francis Bach. In 2018, he received the prize of the Fondation Mathematique Jacques Hadamard for the best PhD thesis in the field of optimization. His research focuses primarily on learning problems at the interface of machine learning and optimization.
More information
Practical information
- General public
- Free
- This event is internal
Contact
- Host: Rüdiger Urbanke