Bandeau image

Deep learning toward Artificial Intelligence

Deep learning toward Artificial Intelligence

In the last few years, Deep Learning has achieved quite a few breakthroughs in Data Science, literally yielding performance jumps in computer vision and natural language processing. The reasons for these achievements --besides massive data, computational power, and design efforts-- are not yet fully understood and call for three research challenges.
Corps de texte

The first one is to set up a learning theory suited to analyze deep architectures. The second one concerns their compositionality and their ability to represent higher-order logic models. The third one concerns their interpretation: opening the black box to make the underlying representations explicit.


  • Innovative machine learning and AI: common sense, adaptability, generalization 
  • Deep learning and adversarial learning
  • Machine learning and hyper-optimization
  • Optimization for learning, stochastic gradient method improvements, Bayesian optimization, combinatorial optimization
  • Link between learning and modelling, integration of a priori into learning
  • Repeatability and robust learning
  • Statistical Inference and Validation
  • Composition of deep architectures