Deep Learning and Information Geometry for Time-Series Classification
This work tackles the classification of structured time series, in particular micro-Doppler radar signals from non-cooperative drone targets. We first dwell upon the rich internal structure of such signals, which births a variety of different, yet relatable input representations which in turn may yield a variety of different learning models, namely convolutional and SPD neural networks. We then show how we can adapt known models to the structure of the data through the scope of information geometry, and go further by building novel schemes for the classification of such data. We finally demonstrate a full-classification pipeline making use of all aforementioned structures involved in data formation and classification, and present results thereupon on multiple datasets of real and synthetic data.
Defence : 07/03/2020 - 10h - Visioconférence Jury members : Mme. Florence Tupin, Professeure, Télécom Paris [Rapporteur]
M. Marco Congedo, CR CNRS, Grenoble [Rapporteur]
M. Yannick Berthoumieu, Professeur, IMS Bordeaux
M. Frédéric Barbaresco, Senior scientist, THALES
M. Hichem Sahbi, CR CNRS, Sorbonne Université
M. Olivier Schwander, MdC, Sorbonne Université
M. Matthieu Cord, Professeur, Sorbonne Université
2018-2019 Publications
2019
D. Brooks, O. Schwander, F. Barbaresco, J.‑Y. Schneider, M. Cord : “Riemannian batch normalization for SPD neural networks”, Thirty-third Annual Conference on Neural Information Processing Systems., Vancouver, Canada (2019)
D. Brooks, O. Schwander, F. Barbaresco, J.‑Y. Schneider, M. Cord : “Second-order networks in PyTorch”, GSI 2019 - 4th International Conference on Geometric Science of Information, vol. 11712, Lecture Notes in Computer Science, Toulouse, France, pp. 751-758, (Springer) (2019)
LIP6 » News » PhD students
Deep Learning and Information Geometry for Time-Series Classification