Institut de Mathématiques de Toulouse

Les événements de la journée

2 événements

  • Mathématiques de l’apprentissage

    Jeudi 31 mai 12:30-13:30 - Gilles Richard - IRIT

    The future of ML : autoML !

    Résumé : Forget about ’import keras, model.add(Conv2D(32, (3, 3), input_shape=(64,64,3),activation=’relu’))’,
    too boring… Let’s autoML do the job for you !
    autoML is the last attempt to make computer clever : a proper program can
    get your data, pre-process these data and find the best available predictive model
    for you. In this talk, we will (try to) understand the ideal pipeline from data
    to predictors, and how it can be automatically built.
    We will first see how the issue is theoretically addressed, then practically solved.
    We can consider autoML as a new step toward UML (for Universal Machine Learning !).

    Lieu : Salle 207, bât 1R2

    [En savoir plus]

  • Séminaire Mathématiques pour la biologie

    Jeudi 31 mai 13:30-14:30 - Clément Sire - LPT, IRSAMC

    (Measuring) Social Interactions and (Studying) Collective States in Fish Schools

    Résumé : The flexible coordination of individuals’ movements ensures rapid and coherent changes in direction of travel of fish schools for instance as a reaction to a predator detected in the neighborhood. However the ’microscopic level’ interaction rules involved in the coordination of fish movements and the adapted collective response of a school still remain to a large extent unknown. Knowing such interaction rules could offer new sources of inspiration to design distributed control algorithms for swarms of drones. Here we present a systematic methodology to measure and analyze social interactions controlling the collective motion of animal groups. Contrary to classical forces between physical objects, social interactions between individuals explicitly depend on their relative headings and are affected by their anisotropic and asymmetric perception of their environment. Hence they strongly break the Newtonian’s law of action-reaction. When applied to fish groups, this approach leads to the quantitative measurement of the spontaneous behavior of a fish, of its avoidance interaction with the tank walls, and of its attraction and alignment interaction with another fish. We use the results of this analysis to build an explicit and faithful model that convincingly reproduces quantitative and qualitative features of the actual fish dynamics. We also show that the type of models derived from such analysis reproduces the main collective states observed in actual fish schools, when one varies the intensity of the alignment and attraction interactions between fish.

    [En savoir plus]