Data Science Seminar

Le 29 novembre 2018 à Télécom Paris / Barrault. Séminaire entièrement en anglais.
The LTCI Data Science Seminar is a joint research seminar between the DIG and the S2A teams. It focuses on machine learning and data science topics.
November 29, 2018

The seminar took place from 2PM to 4PM (room C48), and featured two talks:

Talk 1: Rodrigo Mello (University of Sao Paulo): Steps, concepts and issues involved in providing learning guarantees in the Deep Learning scenario (most specially Convolutional Neural Networks)

You can download the slides of this talk.

Abstract: The main idea about this one-hour talk is to start with the Generalization bound, from the Statistical Learning Theory, estimate the Shattering coefficient of a single-hyperplane-based algorithm in different input space dimensionalities, so that we end up questioning the complexity of Deep Learning biases and how they could be estimated.

Talk 2: Olivier Sigaud (Sorbonne University): Efficient exploration for learning continuous action policies

You can download the slides of this talk.

Abstract: This talk is based on our ICLM 2018 paper with Cédric Colas and Pierre-Yves Oudeyer. I will first give a tutorial introduction to reinforcement learning (RL) and the DQN and DDPG deep RL algorithms. Then I will highlight a simple exploration issue faced by these systems, I will present Goal Exploration Processes (GEPs) as an efficient exploration method, and I will propose a way of combining GEPs with DDPG which provides a satisfactory solution to the issue. Finally, I will wrap these results into the broader context of Artificial Intelligence from a developmental learning perspective.