Prof. Stéphane Mallat: Modeling Deep Networks: Network Learning for Image Processing

Event details
Date | 18.11.2021 |
Hour | 17:00 › 18:00 |
Speaker | Stéphane Mallat, Collège de France |
Location | Online |
Category | Conferences - Seminars |
Event Language | English |
This event is part of the EPFL Seminar Series in Imaging
Abstract. Deep neural network learning from data has taken over image processing. Not just for classificaiton and regression but also for denoising and inverse problems. Is it the end of geometric models and understanding ? Deep network models are high-dimensional and must be analyzed in a probabilistic framework. Yet they must also take into account image properties, including multiscale structures and symmetries. The lecture takes an information theory point of view, and shows that the underlying mathematics are closely related to statistical physics. We introduce a general class of interpretable neural network models through the renormalisation group and multiscale wavelet transforms, with applications to image generation and classification.
Biography. Stéphane Mallat’s research interests include machine learning, signal processing, and harmonic analysis. Starting from truly original theoretical work, he has developed their applications up to industrial transfer, with 10 international patents. He holds a Ph.D. from the University of Pennsylvania. Since 2017, he has held the “Data Sciences” chair at the Collège de France. From 2001 to 2007 he was co-founder and CEO of a semiconductor start-up company which has grown into a semiconductor company manufacturing millions of electronic chips to increase the resolution of pictures in high definition televisions. He is a member of the French Academy of sciences, a foreign member of the US National Academy of Engineering, an IEEE Fellow and an EUSIPCO Fellow.
Abstract. Deep neural network learning from data has taken over image processing. Not just for classificaiton and regression but also for denoising and inverse problems. Is it the end of geometric models and understanding ? Deep network models are high-dimensional and must be analyzed in a probabilistic framework. Yet they must also take into account image properties, including multiscale structures and symmetries. The lecture takes an information theory point of view, and shows that the underlying mathematics are closely related to statistical physics. We introduce a general class of interpretable neural network models through the renormalisation group and multiscale wavelet transforms, with applications to image generation and classification.
Biography. Stéphane Mallat’s research interests include machine learning, signal processing, and harmonic analysis. Starting from truly original theoretical work, he has developed their applications up to industrial transfer, with 10 international patents. He holds a Ph.D. from the University of Pennsylvania. Since 2017, he has held the “Data Sciences” chair at the Collège de France. From 2001 to 2007 he was co-founder and CEO of a semiconductor start-up company which has grown into a semiconductor company manufacturing millions of electronic chips to increase the resolution of pictures in high definition televisions. He is a member of the French Academy of sciences, a foreign member of the US National Academy of Engineering, an IEEE Fellow and an EUSIPCO Fellow.
Practical information
- Informed public
- Free