Mystères mathématiques de l’apprentissage par réseaux de neurones

I wanted to recommend this talk by Stéphane Mallat (en Français), in which he tries to propose a way of interpreting and understanding deep CNNs using in a sort of wavelet network. The first 30 minutes are pretty much a review of what we know on deep learning (of course oriented to what he presents next). Nevertheless I appreciated his point of view and did gain some insights from it.
Video: https://www.youtube.com/watch?v=xqeuamn_o_I
Related paper: http://rsta.royalsocietypublishing.org/content/374/2065/20150203

Disclaimer: originally posted internally at Mnemosyne’s Openlab