Research interests

The central focus of the team is Statistical Machine Learning and Deep learning. We develop algorithmic and theoretical contributions together with applications in selected domains.

Algorithmic and theoritical contributions

Representation learning and Deep learning

Representation learning emerged over the last ten years at the confluence of different domains with the aim of learning meaningful representations from data. The flagship today is represented by Deep Neural Networks. MLIA has a strong historical position on this domain and a large part of its research is within this field. Over this period, we have investigated learning representation for structured and dynamic data, with applications in Computer Vision, Natural Language Processing, Social Data Analysis and Recommendation.

Structured outputs

Learning to produce complex outputs encompasses many fundamental and application problems. A research direction has been Extreme Classification, a new challenge consisting in classifying with millions of classes. MLIA has organized 4 international challenges and workshops on this topic and participated to a EU project. We are also investigating classification in graphs for social networks and image segmentation. In a Franco-Canadian project, we are combining deep learning and structured prediction.

Sequential and Reinforcement learning

Many machine learning problems can be revisited as sequential learning problems. This allows one to define new learning methods, and to extend existing models to more complex problems. We have investigated the development of new sequential models for different key tasks like structured output prediction, graph processing, budgeted learning and attention models. We have organized one workshop on this topic at ICML 2013.

Application domains

Computer vision

We have developed several models for visual pattern detection and recognition. We have explored together computer vision-based approaches, bio-inspired modeling, and deep learning strategies for many years. This gives to MLIA a strong position in the current era of deep convolutional neural nets.

Natural Language Processing and Information Retrieval, Recommendation

MLIA has been involved in the text community for years with contribution on learning to rank and on semi-structured data retrieval. We have developed models for joint review, text polarity and recommendation prediction. Recently,  we investigate a new line of research on Deep Learning for IR and natural language understanding.

Complex dynamic data analysis

We have developed statistical models for the analysis and modeling of information diffusion on social net by formulating the problem in continuous spaces instead of discrete ones. We have also developed a cooperation with actors like Renault and Ile de France Transportations (STIF) for analyzing problems in the transportation domain.

 

Future work:

  • Multi-modal perception
  • Unsupervised learning and learning with weak supervision
  • Statistico-mecanical models and spatio-temporal processing
  • Human-Machine interaction