Workshop Optimization and Learning

Optimization and Learning

CIMI Workshop, Toulouse, France

10-13 september 2018


The « Optimization and Learning » workshop will take place at Institut de Mathématiques de Toulouse (IMT) and is part of the thematic semester « Optimization » organized by Labex CIMI.

The workshop will focus on important challenges in optimization for machine learning, pertaining to the following subtopics:

  1. Representation learning.  Artificial intelligence, signal & image processing and automatic language processing, among other disciplines, motivate the development of novel methodology for matrix factorization, dictionary learning and deep learning: the workshop will present recent results on the characterization and obtention of the solutions to these non-convex problems.

  2. Stochastic optimization.  In some problems pertaining to computational statistics, signal & image processing, risk computations or learning in large dimension, the objective function is intractable or available up to some approximations. The workshop will present methodological & theoretical advances to address such settings, in the context of non-smooth and non-convex optimization and including online and distributed algorithms.

  3. Optimization with uncertainty. The workshop will address the optimization of backbox functions, possibly in the presence of noise, by means of Gaussian Processes (GP) or bandit models. It will include a mini-course on the use of GPs for optimization, and a series of presentations on recent advances in multi-armed bandit models and GP-based procedures, with a focus on theoretical guarantees and in particular bounds on the cumulated regret.


The workshop will include three mini-courses (3 x 3h), a dozen one hour talks, and poster sessions.

Speakers for the mini-courses (3 courses of 3h each)

Speakers for the one hour talks

Participants are invited to present their work during a poster session (see the registration form for the application).



The workshop will start on Monday September 10, at 11:00 a.m. and will end on Thursday September 13 at 4:00 p.m. It will take place in Amphi Schwarz, Building 1R3, Institut de Mathématiques de Toulouse (IMT). Click here for a Google Map link.

Detailed program with abstracts can be dowloaded here.

Registration desk opens on Monday September 10th at 10:30am.

Monday September, the 10th
11:00 - 11:30: Welcome coffee 11:30 - 13:00: G. Lan -- Stochastic Optimization Algorithms for Machine Learning (Mini Course 1/2). 13:00 - 14:30: Lunch
14:30 - 15:30: E. Gobet -- Uncertainty Quantification of Stochastic Approximation Limits.
15:30 - 16:30: C. Szepesvari -- Completing the classification of adversarial partial monitoring games.
16:30 - 17:00: Coffee break
17:00 - 18:00: L. Zdeborova -- Constrained low-rank matrix estimation.

Tuesday September, the 11th
09:00 - 10:30: G. Lan -- Stochastic Optimization Algorithms for Machine Learning (Mini Course 2/2).
10:30 - 11:00: Coffee break
11:00 - 12:00: G. Biau -- Some theoretical properties of GANs.
12:00 - 13:00: J. Lee -- Geometry of Optimization Landscapes and Implicit Regularization of Optimization Algorithms.
13:00 - 14:00: Lunch
14:00 - 14:30: Poster session 1
14:30 - 16:00: R. Vidal -- Global Optimality in Matrix Factorization, Tensor Factorization and Deep Learning (Mini Course 1/2).
16:00 - 16:30: Coffee break
16:30 - 17:30: T. van Erven -- MetaGrad: Multiple learning rates in online learning.
17:30 - 18:30: M. Valko -- Active block-matrix completion with adaptive confidence sets.

Wednesday September, the 12th
09:00 - 10:30: R. Vidal -- Global Optimality in Matrix Factorization, Tensor Factorization and Deep Learning (Mini Course 2/2).
10:30 - 11:00: Coffee break
11:00 - 12:00: O. Shamir -- Optimization Landscape of Neural Networks: Where Do the Local Minima Hide?
12:00 - 13:00: N. Gillis -- Computing nonnegative matrix factorizations.
13:00 - 14:00: Lunch
14:00 - 14:30: Poster session 2
14:30 - 16:00: A. Krause -- Bayesian optimisation and Gaussian process bandits: Theory and Applications (Mini Course 1/2).
16:00 - 16:30: Coffee break
16:30 - 17:30: F. Bach -- Statistical Optimality of Stochastic Gradient Descent on Hard Learning Problems through Multiple Passes.
17:30 - 18:30: A. Iouditsky -- Estimate aggregation from indirect observations.

Thursday September, the 13th
09:00 - 10:30: A. Krause -- Bayesian optimisation and Gaussian process bandits: Theory and Applications (Mini Course 2/2).
10:30 - 11:00: Coffee break
11:00 - 12:00: C. Couprie -- Future video prediction and creative image generation.
12:00 - 13:00: C. Gentile -- Nonstochastic Bandit Bros: Vanilla, Partial, Delayed, Composite, Contextual.
13:00 - 14:30: Lunch
14:30 - 15:30: R. Combes -- Minimal Exploration in Structured Stochastic Bandits.
15:30 - 16:30: F. Panloup -- Non asymptotic analysis of the Ruppert-Polyak averaging algorithm.




Tuesday September, the 11th from 2:00pm to 2:30 pm

Abdessamad Amir: Newton method with an adjusted generalized Hessian matrix for SVMs.
Averyanov Yaroslav: Early stopping rule and discrepancy principle in RKHS
Besson Lilian: Multi-Player Bandits Revisited Bittar Thomas: Kriging techniques and decomposition method to optimize a risk criterion on the Net Present Value
Cambareri Valerio: TBA
Debarnot Valentin: A scalable estimator of sets of integral operators
Kervazo Christophe: Heuristics for Efficient Sparse Blind Source Separation
Koolen Wouter M.: TBA   
Leglaive Simon: A variance modeling framework based on variational autoencoders for speech enhancement
Locatelli Andrea: Adaptivity to smoothness in nonparametric optimization. -- slides

Wednesday September, the 12h from 2:00pm to 2:30 pm

Barbaresco Frederic: Information Geometry & Fisher-Koszul-Souriau metric and their uses in Machine Learning
Besson Rémi: A model-based reinforcement learning approach for a rare disease diagnostic task
Logé Frédéric: Revisiting the greedy algorithm for Contextual Bandits.
Fagot Dylan; Nonnegative Matrix Factorization With Transform Learning
Filstroff Louis: Closed-form Marginal Likelihood in Gamma-Poisson Matrix Factorization
Gower Robert: Tracking the gradients using the Hessian: A new look at variance reducing stochastic methods
Ostrovskii Dmitrii: Non-asymptotic Analysis of M-estimators via Self-concordance
Priem Rémy: Super Efficient Global Optimization with Mixture of Experts
Vono Maxime: Split-and-augmented Gibbs sampler - A divide & conquer approach to solve large-scale inference problems



Registration is free but mandatory. The participants are invited to present a poster: A title and a short abstract can be given when filling the registration form.

Registrations are closed.




Organizing commitee:

François Bachoc, Cédric Févotte, Gersende Fort, Sébastien Gadat and Laurent Risser

Advisory committee:

Nicolas Dobigeon, Aurélien Garivier, François Malgouyres, Edouard Pauwels and Aude Rondepierre




For any question, please contact Sébastien Gadat: firstname.lastname(no accents)




logocimi     logo_iuf_crop     logo_erc     logo-iuct