Active Learning and Optimization
Experimental design as a support to heavy computer experiments or to complex system design can be seen as the problem of optimizing an unknwown function, generally lacking of ‘nice’ properties beyond regularity and for which each evaluation comes with a cost. The framework for prediction and decision involves sequential resource allocation and selective sampling. MLMDA team has proposed various methodologies and algorithms that explore old questions such as experimental design and inverse problems with modern machine learning algorithms.
Machine Learning on Large Networks
Understanding the dynamics of information cascades and epidemics on networks involves advanced modeling of diffusion processes on graph structures. MLMDA team studies the conditions regulating the behavior of the network confronted to such propagation phenomena, as well as resource allocation strategies that permit to contain and eventually to control the contagion. The contributions are methodological, theoretical and computational.
Machine Learning and Signal Processing on Physiological Signals
Posture, walk, ocular movement can now be recorded with a variety of low-cost sensors. Jointly with neurophysiologists and clinicians from COGNAC G (CNRS & Université Paris Descartes), MLMDA team contributes to the design of complete and replicable protocols for data collection, and digital processing methods for event detection based on multichannel sensor signals. The algorihms developed cover both low-level signal processing to statistical detection, learning and prediction.