Journées de l'optimisation 2022

HEC Montréal, Québec, Canada, 16 — 18 mai 2022

Horaire Auteurs Mon horaire

WB7 - Optimization and Learning II

18 mai 2022 13h30 – 15h10

Salle: Quebecor (jaune)

Présidée par Dounia Lakhmiri

4 présentations

  • 13h30 - 13h55

    Optimization for Federated Learning

    • Guojun Zhang, prés., Huawei Noah's Ark Lab

    With the wide deployment of machine learning (ML) systems, it has been a critical problem to train ML models while keeping the privacy of user data. Federated learning (FL) is such a scheme that aims to solve the privacy issue. The optimization of FL can be traced back to the established field of distributed optimization, but is more challenging due to the heterogeneity of various clients. In this tutorial, I will introduce federated optimization, including the problem setting, the main challenges as well as different recent approaches to federated optimization. I will conclude this talk by highlighting some unresolved problems that are important for future research.

  • 13h55 - 14h20

    The role of AI in enhancing students learning experience based on the Andragogy learning model: The perspectives in Industrial engineering courses

    • Muhammad Azam, Polytechnique Montreal
    • Ahmad Shahnejat Bushehri, prés., Polytechnique Montreal
    • Samira Keivanpour, Polytechnique Montréal

    Artificial Intelligence (AI) has changed the paradigm in education and training. The Andragogy model is one of the best learning models in the active pedagogical context. The principles of the model include involving the students in the planning and evaluation phase, learning by doing for an active learning experience, and real-life problem-based learning. We have adopted this model in the context of teaching industrial engineering courses and presented the contribution of AI in facilitating the elements of this model. A brief literature review is provided to show the integration of clustering, data mining, and learning-based algorithms in the educational context. The examples in the production/manufacturing and sustainable manufacturing courses are provided to show the implications on the quality of teaching and students' learning experience.

  • 14h20 - 14h45

    Copula transfer learning for population synthesis

    • Pascal Jutras-Dubé, prés., University of Montreal
    • Fabian Bastin, University of Montreal
    • Cinzia Cirillo, University of Maryland
    • Javier Bas, University of Maryland
    • Mohammad Bilal Mohammad Al-Khasawneh, University of Maryland
    • Zhichao Yang, University of Maryland

    Population synthesis consists of generating synthetic but realistic representations of a target population of micro-agents for the purpose of behavioral modeling and simulation. Traditional deep generative modeling approaches are well suited to generate a synthetic population from a given sample, but do not explicitly incorporate observed empirical marginal distributions. We introduce a framework that can generate synthetic data for a target population of which only the empirical marginal distributions are known by using a sample from another population sharing similar marginal dependencies. Specifically, we normalize the data with empirical copulas to separate the learning of the marginal distributions from the dependence structure of the target population, which allows the injection of marginal information into a traditional population generation approach based on machine learning. We illustrate on American Community Survey (ACS) data that the proposed method makes it possible to study the structure of the data in a way that is robust to the peculiarities of the marginal distributions. We successfully generate a synthetic sample of a target population from which we have information on the marginals only by using a sample from another source population sharing the dependencies structure of the target population.

  • 14h45 - 15h10

    A Stochastic Proximal Method for Nonsmooth Regularized Finite Sum Optimization

    • Dounia Lakhmiri, prés., Polytechnique Montreal
    • Dominique Orban, GERAD - Polytechnique Montréal
    • Andrea Lodi, University of Bologna

    We consider the problem of training a deep neural network with nonsmooth regularization to retrieve a sparse and efficient substructure. Our regularizer is only assumed to be lower semicontinuous and prox-bounded. We combine an adaptive quadratic regularization approach with proximal stochastic gradient principles to derive a new solver, called SR2, whose convergence and worst-case complexity do not require knowledge or approximation of the gradient's Lipschitz constant. Our experiments on network instances trained on CIFAR-10 and CIFAR-100 with $\ell_1$ and $\ell_0$ regularization show that SR2 achieves competitive accuracy and sparsity levels.