Il ne fallait pas moins que les locaux ultras design de Criteo pour accueillir cette troisième édition du Paris Meetup Deep Learning, co-organisée avec le Paris Meetup Machine Learning. 160 participants et toujours des interventions de très grandes qualités !
Retrouver toute la session en vidéo ici.
Rappel du programme :
- Yoshua Bengio : Deep Learning Theory (remote from London – London Machine Learning meetup). Although neural networks have long been considered lacking in theory and much remains to be done, theoretical evidence is mounting and will be discussed, to support distributed representations, depth of representation, the non-convexity of the training objective, and the probabilistic interpretation of learning algorithms (especially of the auto-encoder type, which were lacking one). The talk will focus on the intuitions behind these theoretical results.
- Sander Dieleman, http://benanne.github.io/ : classifying plankton with deep neural networks by the Deep Sea team from Reservoir Lab – Ghent University. Deep learning has become a very popular approach for solving computer vision problems in recent years. In this talk we’ll demonstrate how this approach can be applied in practice. We’ll show how our team of 7 built a model for the automated classification of plankton based on convolutional neural networks. Using this model, we placed 1st in the National Data Science Bowl competition on Kaggle.
- Gabriella Contardo, LIP6, UPMC : learning to build representations from partial information – Application to cold-start recommendation http://www.lip6.fr/actualite/personnes-fiche.php?ident=D1412. Most of the successful machine learning algorithms rely on data representation, i.e a way to disentangle and extract useful information from data, which will help the model in its objective task. Classical approaches build representations based on fully observed data. But in many cases, one wants to build representations »on the fly », based on a partially observed information. As an example, learning representations over users can be done by progressively gathering information about their profiles. This paper presents an inductive representation-based model to tackle the twofold more general problem of (i) selecting the right information to collect for building relevant representations, (ii) updating these representations based on new incoming information. It is developed in this paper to design static interview for the cold-start collaborative filtering problem but it can also be used to go smoothly to the warm context where all information has been gathered.
- Guillaume Wenzek : analyse de sentiment à l’aide de réseaux de neurones récursifs / Sentiment Analysis With Recursive Neural Tensor Network. Sentiment analysis is one of the hardest NLP (Natural Language Processing) task, due to complex linguistic structures such as negation or double-negation. Socher et al. introduced a method that combines a classic NLP tool, a syntaxic parser, with a special kind of neural networks. We will review this method and introduce a few improvements in order to train on a corpus with fewer annotations than the Stanford Sentiment Treebank used in the paper.