Научны(е)й руководител(и)ь : Matthieu CORD

Со-руководитель : THOME Nicolas

Problématiques de modélisation et d'apprentissage par des architectures profondes en vision par ordinateur

Study on training methods and generalization performance of deep learning for image classification.
Artificial intelligence has gained a lot of interest in recent years, particularly with the impressive advances in computer vision. More specifically, deep neural networks today enable us to automate many tasks in the processing of visual information. This family of machine learning algorithms demonstrates a remarkable facility to learn large datasets. Despite their impressive performance, their ability to generalize remains largely misunderstood. From a research and application point of view, deep learning performance and explicability are more and more demanding. This is the purpose of our research whose contributions are presented in this thesis.
We first worked on accelerating the training of deep networks via distributed computing methods among several GPUs. Decreasing the optimization time of neural networks makes it easier to test new architectures as well as new learning methods. We then studied the convolution neural network architectures in order to improve them, without increasing their complexity. We have proposed a new activation function, which filters less information than the standard ReLU function, allowing better processing of semantic information. Finally, we considered regularizing the training of networks. In particular, we have studied a criterion of regularization based on information theory, which we have deployed in two different ways. The first instantiation models the criterion layer by layer for more efficient regularization. The second uses a variational to bound the criterion and optimize it. Both methods have been rigorously validated by comparison with the state of the art on many image classification reference databases, such as ImageNet.

Публикации 2016-2019

• 2019
  • A. Saporta, Y. Chen, M. Blot, M. Cord : “REVE: Regularizing Deep Learning with Variational Entropy Bound”, 2019 IEEE International Conference on Image Processing (ICIP), Taipei, Taiwan, Province of China, pp. 1610-1614, (IEEE) (2019)
  • M. Blot, D. Picard, N. Thome, M. Cord : “Distributed Optimization for Deep Learning with Gossip Exchange”, Neurocomputing, vol. 330, pp. 287-296, (Elsevier) (2019)
• 2018
  • M. Blot : “Problématiques de modélisation et d’apprentissage par des architectures profondes en vision par ordinateur”, диссертация, Защита диссертаций 19.11.2018, Научны(е)й руководител(и)ь Cord, Matthieu, Со-руководитель : Thome, Nicolas (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)
• 2016
  • M. Blot, M.‑J. Lesot, M. Detyniecki : “Transformation-based constraint-guided Generalised Modus Ponens”, Proc. of the IEEE WCCI International Conference, Vancouver, Canada (2016)
  • M. Blot, M. Cord, N. Thome : “Max-min convolutional neural networks for image classification”, Image Processing (ICIP), 2016 IEEE International Conference on, Phoenix, United States, pp. 3678-3682, (IEEE) (2016)