رئاسـة البـحث : Matthieu CORD
تأطـير مـشـترك : THOME Nicolas

Improving ConvNets Latent Representations for Visual Understanding

For a decade now, convolutional deep neural networks have demonstrated their ability to produce excellent results for computer vision. For this, these models transform the input image into a series of latent representations. In this thesis, we work on improving the ``quality'' of the latent representations of ConvNets for different tasks. First, we work on regularizing those representations to increase their robustness toward intra-class variations and thus improve their performance for classification. To do so, we develop a loss based on information theory metrics to decrease the entropy conditionally to the class. Then, we propose to structure the information in two complementary latent spaces, solving a conflict between the invariance of the representations and the reconstruction task. This structure allows to release the constraint posed by classical architecture, allowing to obtain better results in the context of semi-supervised learning. Finally, we address the problem of disentangling, i.e. explicitly separating and representing independent factors of variation of the dataset. We pursue our work on structuring the latent spaces and use adversarial costs to ensure an effective separation of the information. This allows to improve the quality of the representations and allows semantic image editing.

مناقـشـة مـذكـرة : 03/10/2019

أعـضاء لجنة المناقـشة :

Stéphane Canu, INSA Rouen / LITIS [rapporteur]
Greg Mori, Simon Fraser University & Borealis AI [rapporteur]
Catherine Achard, Sorbonne Université / ISIR
Kartheek Alahari, Inria Grenoble / Thoth
David Picard, École nationale des ponts et chaussées / IMAGINE
Nicolas Thome, CNAM / CEDRIC
Matthieu Cord, Sorbonne Université / LIP6 & Valeo.ai

إصدارات 2010-2021

• 2021
  • A. Douillard, E. Valle, Ch. Ollion, Th. Robert, M. Cord : “Insights from the Future for Continual Learning”, CVPR Workshop, Nashville, United States, pp. 3477-3486, (IEEE) (2021)
• 2020
  • A. Douillard, M. Cord, Ch. Ollion, Th. Robert, E. Valle : “PODNet: Pooled Outputs Distillation for Small-Tasks Incremental Learning”, ECCV 2020 - 16^th European Conference on Computer Vision, vol. 12365, Lecture Notes in Computer Science, Glasgow, United Kingdom, pp. 86-102, (Springer) (2020)
• 2019
  • Th. Robert : “Amélioration des représentations latentes des ConvNets pour l’interprétation de données visuelles”, أطروحة, مناقـشـة مـذكـرة 03/10/2019, رئاسـة البـحث Cord, Matthieu, تأطـير مـشـترك : Thome, Nicolas (2019)
• 2018
  • M. Blot, Th. Robert, N. Thome, M. Cord : “SHADE: Information-Based Regularization for Deep Learning”, ICIP 2018 - 25^th IEEE International Conference on Image Processing, Athènes, Greece, pp. 813-817, (IEEE) (2018)
  • Th. Robert, N. Thome, M. Cord : “HybridNet: Classification and Reconstruction Cooperation for Semi-supervised Learning”, Computer Vision – ECCV 2018 15^th European Conference, Munich, Germany, September 8–14, 2018, Proceedings, vol. 11211, Lecture Notes in Computer Science, Munich, Germany, pp. 158-175, (Springer) (2018)
• 2010
  • G. Lasnier, Th. Robert, L. Pautet, F. Kordon : “Behavioral Modular Description of Fault Tolerant Distributed Systems with AADL Behavioral Annex”, 10^th international conference on New Technologies of Distributed Systems (NOTERE'2010), Tozeur, Tunisia, pp. 17-24, (IEEE) (2010)
  • G. Lasnier, Th. Robert, L. Pautet, F. Kordon : “Architectural and Behavioral Modeling with AADL for Fault Tolerant Embedded Systems”, 13^th IEEE International Symposium on Object-oriented Real-time distributed Computing (ISORC'10), Carmona, Spain, pp. 87-91, (IEEE) (2010)