Asymptotics and Dynamics of Learning,
Event details
| Date | 16.06.2023 |
| Hour | 13:00 › 15:00 |
| Speaker | Yatin Dandi |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Nicolas Flammarion
Thesis advisor: Prof. Lenka Zdeborova
Co-examiner: Prof. Lénaïc Chizat
Abstract
Theoretical analysis of the asymptotics of systems involving a large number of interacting entities has enabled massive progress in fields ranging from statistical physics, high dimensional statistics, combinatorial optimization, to the theory of machine learning. In this PhD, we aim to build upon the techniques developed in the above fields to obtain novel insights into the phase transitions and dynamics in high-dimensional computational phenomenon, particularly for domains such as representation learning and causality, where many central questions remain unanswered.
Background papers
High-dimensional Asymptotics of Feature Learning: How One Gradient Step Improves the Representation: https://openreview.net/pdf?id=akddwRG6EGi
Online stochastic gradient descent on non-convex losses from high-dimensional inference: https://jmlr.csail.mit.edu/papers/volume22/20-1288/20-1288.pdf
Majority dynamics on trees and the dynamic cavity method: https://arxiv.org/abs/0907.0449, https://projecteuclid.org/journals/annals-of-applied-probability/volume-21/issue-5/Majority-dynamics-on-trees-and-the-dynamic-cavity-method/10.1214/10-AAP729.full
Exam president: Prof. Nicolas Flammarion
Thesis advisor: Prof. Lenka Zdeborova
Co-examiner: Prof. Lénaïc Chizat
Abstract
Theoretical analysis of the asymptotics of systems involving a large number of interacting entities has enabled massive progress in fields ranging from statistical physics, high dimensional statistics, combinatorial optimization, to the theory of machine learning. In this PhD, we aim to build upon the techniques developed in the above fields to obtain novel insights into the phase transitions and dynamics in high-dimensional computational phenomenon, particularly for domains such as representation learning and causality, where many central questions remain unanswered.
Background papers
High-dimensional Asymptotics of Feature Learning: How One Gradient Step Improves the Representation: https://openreview.net/pdf?id=akddwRG6EGi
Online stochastic gradient descent on non-convex losses from high-dimensional inference: https://jmlr.csail.mit.edu/papers/volume22/20-1288/20-1288.pdf
Majority dynamics on trees and the dynamic cavity method: https://arxiv.org/abs/0907.0449, https://projecteuclid.org/journals/annals-of-applied-probability/volume-21/issue-5/Majority-dynamics-on-trees-and-the-dynamic-cavity-method/10.1214/10-AAP729.full
Practical information
- General public
- Free