Separation/extraction of sources are wide concepts in information sciences, since sensors provide information mixing and an essential step consists in separating or extracting useful information from unuseful one, called noise. In this project, we consider three challenges.The first one is the multimodality. Indeed, with the multiplication of kinds of sensors, in many areas like biomedical signal processing, hyperspectral imaging, etc. there are many ways for recording the same physical phenomenon leading thus to multimodal data. Multimodality has been studied in the framework of human-computer interface or in data fusion, but never at the signal level. The objective is to provide a general framework for modeling classical multimodal properties, like complementarity, redundancy, equivalence, etc. as of function of source signals.The second challenge is nonlinearity. Indeed, there exist a few cases where the mixtures are essentially nonlinear, e.g. with chemical sensors. The main objective is to enlarge results on identifiability conditions for new classes of nonlinearities and priors on sources.The third challenge is the data size. For high-dimension data (e.g. EEG or MRI in brain imaging), separating all the sources is neither tractable nor relevant, since one would like to only extract the useful sources. Conversely, for a small number of sensors, especially smaller than the number of sources, it is again necessary to only focus on the useful signals. The main objective is to develop generic approaches able to only extract useful signals, based on simple reference signal, modeling weak properties of the useful signal.Finally, validation and relevant modeling must be based on actual signals and problems. In this project, theoretical results and algorithms will be developed in interaction with applications in biomedical engineering (brain-computer interface, EEG, fMRI), chemical engineering, audio-visual scene analysis and hyperspectral imaging.