Seminars

Overview

The Thoth seminar series is a weekly seminar of the Thoth team at Inria Grenoble Rhones-Alpes. Talks are usually at 11am on Thursday. We combine internal speakers from the group with invited speakers, most commonly presenting their recent research but occasionally a broader survey of a topic.
Talks cover a variety of topics in machine learning.
This group is open to everyone. Please contact the organizers if you would like to join our mailing list, to which we send the link for each online talk.

Contact:
Michael Arbel: michael.arbel@inria.fr

Upcoming seminars

  • TBA
    Date: September 23, 2021, 11:00 am

    TBA

  • TBA
    Date: September 30, 2021, 11:00 am

    TBA

  • TBA
    Date: October 7, 2021, 11:00 am

    TBA

Previous seminars

  • Learning Temporal Dynamics for Video Action Recognition
    Date: September 16, 2021, 11:00 am

    Location: Room F107, Centre Inria Grenoble - Rhône-Alpes

    Speaker: Heeseung Kwon - Affiliation: Thoth team - Inria Grenoble - Rhône-Alpes

    Abstract: In this talk, we will consider the task of video action recognition, classifying a human action from a short video clip (~10s). Since a video consists of hundreds of frames, there are two major challenges in video action recognition; learning temporal dynamics in a video & handling the high computational cost of processing the video. The goal of this talk is to show how to design neural architectures for learning temporal dynamics while consuming a less amount of computational cost. Starting from a lightweight neural block that learns motion between two consecutive frames, we will then show an advanced neural block that learns general temporal dynamics in a video. Lastly, we introduce a novel operation for temporal modeling, which can be used as a basic brick (e.g. convolution, self-attention) for video processing architectures.

  • Dual approaches for decentralized optimization in graphs
    Date: September 9, 2021, 11:00 am

    Location: Room F107, Centre Inria Grenoble - Rhône-Alpes

    Speaker: Hadrien Hendrikx - Affiliation: Sierra team - Inria Paris

    Abstract: In this talk, we will consider the problem of minimizing a function f = sum_i^n f_i, where we assume that each summand f_i is held by a node of the network, which can compute the value or the gradient of this function. This corresponds to the Empirical Risk Minimization problem with a distributed dataset, and generalizes the classical gossip averaging problem. We will consider problems in which nodes are linked by a specific communication graph, and can only communicate with their neighbours in this graph. The goal of this talk is to show how to design optimization algorithms in this context and to highlight the interplay between relevant quantities from optimization (e.g., condition number) and graph theory (e.g., spectral gap of the graph Laplacian) in the convergence speed of said algorithms. Starting from a very simple algorithm, we will then show how the dual approach allows us to easily extend powerful tools from convex optimization, e.g., acceleration and variance reduction, to the decentralized setting.

All seminars

Events in September 2021–February 2022

  • -
    Dual approaches for decentralized optimization in graphs
  • -
    Learning Temporal Dynamics for Video Action Recognition
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA
  • -
    TBA

Comments are closed.