Bak to the general presentation

Machine Learning is the branch of computer science that aims at reproducing the learning capabilities of humans, with algorithms that are implemented as computer programs. The methods developed in the framework of machine kearning proved extremely powerful for designing predctive models of the behaviour of complex processes: when domain knowledge is unavailable, or too inaccurate, predictive models can be inferred from examples by machine learning algorithms.

PRESENT PROJECTS

• Learning from structured data: appplication to drug discovery.
• Visuo-acoustic modeling by machine learning: application to the design of a "silent phone".
• Modeling 1D- or 2D-signals by machine learning techniques: application to computer-aided diagnostic from electrophysiological recordings of cardiac or brain activity.
• Dynamic modeling by support vector machines: appplication to the control of pollutants from vehicles.
• Design of virtual sensors by machine learning: applications to environment and health.

Learning from structured data: application to drug discovery

"Graph Machines" were developed in our laboratory, on porder to learn from data structured as graphs, while conventional learning machines learn from data represented by vectors. The method is especially suitable for drug design, because of the graph structure inherent to molecules. The advantages of "graph machines" over concventional QSAR/QSPR (Quantitative Structure-Activity Relations/Quantitative Structure-Property Relations) has been demonstrated on numerous examples.

Visuo-acoustic modeling by machine learning: application to the design of a "silent phone"

The OUISPER project (Oral Ultrasound Synthetic Speech Source), supported by ANR and DGA, aims at designing a system that synthesizes an intelligible speech signal from ultrasound images of the tongue and video images of the lips, withous a need of vocalizing. Two types of applications are currently under consideration:

• protheses for patients without, or with imaired, vocal chords
• a silent phone for use in situations where silence is mandatory

Modeling 1D- or 2D- signals by machine learning techniques: application to computer-aided diagnostic from electrophysiologicval recordings of cardiac or brain activity

Electrophysiological signals are frequently in the form of "bumps". Machine learning allows the modeling of such signals ("bump modeling") in such a way that the model is readily understandable by the physician.

• Electro-cardiographic signals exhbit a characteristic structure of positive and negative peaks (P, Q, R, S and T "waves") that arise from polarozation and depolarization phenomonea that are clearly identified. Pathologies can be detected from alterations in shape, amplitude and position of those peaks. In long-duration recordings (Holter recordings), finding automatically the anomalies in the 100,000 to 150,000 recorded heartbeats is very helpful for the physician. Moreover, computer-aided intra-cardiac surgery is under investigation, in a collaboration with hospitals (Bordeaux, Paris).

• Brain signals such as LFP (Local Field Potential) recordings or electro-encephalographic recordings exhibit characteristic time-frequency structures, which change as a function of the cognitive abilities of the human or animal, especally during training. Bump modeling mde possible the prediction of the development of Alzheimer's disease, 18 months ahead of time ([1], [2]), in patients exhibiting mild cognitive impairments.
  

Dynamic modeling by support vector machines (SVM): application to the control of pollutants from vehicles

Support vector macheines are very powerful methods for automatic classification and regression. The project aims at evaluating the applicability of SVMs to dynamic modeling, especially in coparison with other dynamic modeling techniques such as recurrent neural networks. A preliminary evaluation has been carried out in a collaboration with Renault, in order to control pollutant emissions by vehicles.

Design of virtual sensors by machine learning: applications to environment and health.

Virtual sensors are more and more frequently used in industry: they exempt process designers from installing costly sensors, which are replaced by a predictive model. In a collaboration with unité 551 of INSERM (Pitié-Salpêtrière) and a clinical center, a virtual sensor of strong markers of oxydative stress has been design. It should allow avoidinc costly biological analyses, whose results are predicted from a model whose variables are results of simple analyses (vitamin C, vitamin E, etc.).

OTHER PROJECTS

• Implementing domain knowledge in models designed by machine learing semi-physical modeling.
• Feature selection by the probe feature method ;
• Definition and derivation of a canonical form of discrete-time nonlinear dynamical models;
• Original methods for the initialilzation of wavelet network training;
• An original method for model selection: virtual leave-one-out.

Those results were recognized by numerous French and foreign companies, with which the laboratory carried out research programs that often led to actual applications. :

• Process engineering :
  • semi-physical modeling of a drying process (joint research with 3M USA) ;
  • real-time prediction of the reliability of weldiong spots (joint research with SOLLAC) ;
  • prediction of the mechanical properties of tyre rubber (joint research with MICHELIN) ;
  • modeling of a distillation column for early fault detection (joint research with ELF-ATOCHEM and NETRAL).
• Data mining and language processing : automatic real-time filtering of news releases from Agence France-Presse and automatic generation of information filters ; in that frmaework, our application won the "routing" track in the international competition TREC-9 (Text REtrieval Conference) ;
• Environment monitoring :
  • modeling of sensors in hydrological networks (joint research with Direction de l'Eau et de l'Assainissement of Seine St Denis) ;
  • prediction of ozone pollution peaks.

For more details on those topcis and many others, refer to the publication page. See also our page on the modeling of biological networks.

Click here for information on our former doctoral students.

The modeling of spot welding

Back to top