Complex data, machine learning and representations

Our research team focus on the problems raised by large-scale data management, with a strong orientation towards data whose structure, explicit or not, is complex and requires specific techniques of approximation, extraction and search. The type of data we are dealing with include images, videos, audio or musical documents, satellite imagery and data from multi-spectral sensors. We investigate techniques of statistical machine learning (with a specific focus on deep learning) to extract information, build efficient access techniques and propose new methods of data management based directly on content (as opposed to metadata describing this content). At the moment there are two research axis/directions in our team, described below.

Axis 1. Large image and video databases
We live today in a context characterized by an explosive growth in the production of digital content, doubled by a revolution in digital storage making it possible to keep and easily access large quantities of digital data, beyond the timeline for which it has been initially collected. On the other hand, the rapid development of digital transmission technologies makes possible the distributed distribution and remote sharing of large volumes of such a digital contents. We focus on the structuring, from visual content, of large image and video databases, as well as the search by content in such databases. Our recent work focus on deep learning for the detection of visual patterns and for semantic segmentation of images, the goal being the semantic analysis of scenes taking into account structural and global-local relationships between image components. These approaches also apply very well to data of a different nature, such as musical data, which combine structures at different scales and are generally characterized by a relatively small number of structural items labeled by class.

Axis 2. Music computing and music information retrieval
This axis of research aims to investigate the production of models of musical languages, characterizing homogeneous corpora of music available in symbolic form (scores). Our perspective is to enrich a statistical approach based on explicit data (notes) by a knowledge extraction process identifying the elements of musical language implicitly present in the notation: segmentation in phrases, presence and use of patterns, management of dissonances, cadences, instrumentation and texture. Another direction of research is the development of automatic transcription techniques, conversion of a musical performance to a score in traditional notation by a priori score models (independent of the performance to be transcribed), representing the language of possible musical notations. These techniques can be seen as language models, and are essential components of machine translation or pattern recognition procedures for music data (by analogy with natural language processing).