Skip to content

Learning Theory and Statistical Optimization

Simon Vary edited this page Mar 7, 2025 · 54 revisions

In these meetings, we invite speakers to present on a variety of topics addressing theoretical aspects of machine learning. These meetings are open to members of any department. For people who wish to attend remotely – please contact Simon for access to the Zoom call.

Organisers: Patrick Rebeschini, Varun Kanade, Simon Vary

Schedule

Hilary 2025

This term, the meetings will take place from 16:15 - 17:15 on Thursdays of weeks 7, 8, and 9 in the Meeting Room 2 and the Open Research Area (ORA) of the Department of Statistics.

Date Location Speaker Title
6/3 ORA Alex Buna-Marginean First order methods for linearly separable data
13/3 Meeting Room 2 Sam Howard Schrödinger bridge solvers and connections to mirror descent
20/3 ORA Jack Mayo TBA

Michaelmas 2024

Date Location Speaker Title
23/10 LG.04 Arya Akhavan Blackbox optimization in passive and active schemes
06/11 LG.04 David Janz Why and when randomised exploration works (in linear bandits)
20/11 LG.03 Dorian Baudry Improved learning rates in multi-unit uniform price auctions
27/11 LG.03 Shirley Xiaoqi Liu Loss landscapes and optimization in over-parameterized neural networks: an introduction

Trinity 2024

Date Location Speaker Title
13/06 SLT Johannes Müller Geometry and Convergence of Natural Policy Gradients and Entropic Regularization
04/06 SLT Sílvia Casacuberta Puig Omniprediction: A new paradigm for loss minimization (paper)
21/05 SLT Maria-Alexa Tudose Online and learning-augmented algorithms
30/04 SLT Eugenio Clerico From coin betting to mean estimation and generalisation bounds

Hilary 2024

Date Location Speaker Title
07/03 ORZ Šimon Váry Optimization without retraction on the random generalized Stiefel manifold
29/02 ORZ Peter Potaptchik Diffusion models in infinite dimensions
22/02 SLT Emmeran Johnshon The role of switching cost on minimax rates in online learning
15/02 SLT Amitis Shidani Information Theory, Generative Process and Contrastive Learning

Michaelmas 2023

Date Speaker Title
01/12 Charlie London Generalization Bounds via Convex Analysis (paper)
24/11 Sam Howard Optimal Transport, Sinkhorn's Algorithm and the Entropic Mapping Estimator (paper)
17/11 Carlo Alfano Ordering-based Conditions for Global Convergence of Policy Gradient Methods (paper)
10/11 -- --
03/11 Alex Buna-Marginean First Order Methods through Fenchel Games (paper)

Previous talks

Trinity 2023

Date Speaker Title
05/05 Kaja Gruntkowska Distributed Optimisation with Bidirectional Compression
19/05 Shahine Bouabid Learning with Kernels
16/06 Alireza Amanihamedani Decentralized Online Learning

Hilary 2023

Date Speaker Title
27/01 Alex Buna-Marginean First order methods in robustness
10/02 Group lunch No talk due to COLT deadline
10/03 Emmeran Johnson Minimax lower bounds for policy evaluation in reinforcement learning

Michaelmas 2022

Date Speaker Title
28/10 Carlo Alfano Reinforcement learning: Theory of policy gradient methods
11/11 Eugenio Clerico Transport inequalities and large deviations
25/11 Amitis Shidani Bandit learning: Theory of linear contextual setting
09/12 Tyler Farghly The interface between sampling and optimization: functional inequalities and Langevin Monte Carlo

Hilary 2021

Week Speaker Title
2 Amartya Sanyal Linear Convergence of Gradient and Proximal-Gradient Methods Under the Polyak- Lojasiewicz Condition
3 Gonzalo Mena Statistical Bounds for Entropic Optimal Transport and Sinkhorn EM
4 Eugenio Clerico Some Information-Theoretic Generalization Bounds
5 Eduard Oravkin The Double Descent Phenomenon and Implicit Bias of Mirror Descent
6 David Martinez Accelerating Variance-Reduced Stochastic Gradient Methods
7 --- ---
8 Jun Yang Optimal dimension dependence of the Metropolis-Adjusted Langevin Algorithm

Michaelmas 2020

Week Speaker Title
2 Dominic Richards Learning with Gradient Descent and Weakly Convex Losses (with M. Rabbat)
3 --- ---
4 Valentin De Bortoli Quantitative Propagation of Chaos for SGD in Wide Neural Networks
5 Fan Wu Minimax Rates of Estimation for Sparse PCA in High Dimensions
6 Carlo Alfano On the Theory of Policy Gradient Methods: Optimality, Approximation, and Distribution Shift
7 Tyler Farghly A Non-Asymptotic Analysis of Stochastic Gradient Langevin Dynamics
8 Tomas Vaskevicius Fast Rates and Sparse Linear Prediction

For notes and more information, refer to the old github page.

Clone this wiki locally