Starts 2 Feb 2024 14:00
Ends 2 Feb 2024 16:00
Central European Time
Search
Search in Conferences:
Abdus Salam Distinguished Lecture Series 2024 by Prof. Stéphane Mallat, Collège de France and École normale supérieure, Paris: Lecture 3 "Multiscale neural network models for generation by score diffusion"
ICTP
Leonardo Building - Budinich Lecture Hall
Strada Costiera 11
34151 Trieste
Italy
The 2024 Salam Distinguished Lecture Series will feature theoretical physicist Stephane Mallat.
Stephane Mallat is an applied mathematician, Professor at the College de France. He is a member of the French Academy of Sciences, and a foreign member of the US National Academy of Engineering. He was previously Professor at the Courant Institute of NYU, at Ecole Polytechnique and Ecole Normale Superieure in Paris. He also was a founder and CEO of a semiconductor start-up company. He developed the multiresolution wavelet theory and algorithms at the origin of the compression standard JPEG-2000, and sparse signal representations in dictionaries with matching pursuits. His current work is devoted to mathematical models of deep neural networks, for data analysis and physics.
The Salam Distinguished Lecture Series is an annual presentation of talks by renowned, active scientists. The aim is to showcase important research developments as well as provide a visionary forward view. The lecture series is generously supported by the Kuwait Foundation for the Advancement of Sciences (KFAS).
The overarching topic of the three talks is: Learning Multiscale Energies from Data by Inverse Renormalisation.
Abstract:
Estimating energy models of many body systems amounts to estimate high dimensional probability distributions. It is plagued by the curse of dimensionality. Relying on a inverse renormalization group, we show that this curse is avoided by modeling energy interactions across scales. It leads to well conditioned convex interaction terms that are estimated. Multiscale models are introduced for weak lensing, mass distributions of the cosmic web and turbulence fields, with scattering covariances. They rely on covariances of wavelet coefficient modulus.
Spectacular results have been obtained by generative mdels with damped Langevion diffusions. They rely on energy gradient estimations with deep neural networks that are not well understood. The second colloquim reviews these algorithms. It relates these neural network models to multiscale renormalisation group decompositions. Generalisation and overfitting properties are explained.
There will be 2 colloquium talks and a special topic talk with the following titles:
Colloquium 1: Energy estimation and data generation by inverse renormalisaiton group
Special topic: Log-Sobolev stability, wavelets and interaction energies across scales
Colloquium 2: Multiscale neural network models for generation by score diffusion
The third and last lecture in this Series is on Friday 2 February 2024 starting at 14.00 hrs in the Budinich Lecture Hall and will be on "Multiscale neural network models for generation by score diffusion".
Light refreshments will be served after the talk. All are welcome to attend.
The lectures will also be livestreamed from the ICTP website.
Stephane Mallat is an applied mathematician, Professor at the College de France. He is a member of the French Academy of Sciences, and a foreign member of the US National Academy of Engineering. He was previously Professor at the Courant Institute of NYU, at Ecole Polytechnique and Ecole Normale Superieure in Paris. He also was a founder and CEO of a semiconductor start-up company. He developed the multiresolution wavelet theory and algorithms at the origin of the compression standard JPEG-2000, and sparse signal representations in dictionaries with matching pursuits. His current work is devoted to mathematical models of deep neural networks, for data analysis and physics.
The Salam Distinguished Lecture Series is an annual presentation of talks by renowned, active scientists. The aim is to showcase important research developments as well as provide a visionary forward view. The lecture series is generously supported by the Kuwait Foundation for the Advancement of Sciences (KFAS).
The overarching topic of the three talks is: Learning Multiscale Energies from Data by Inverse Renormalisation.
Abstract:
Estimating energy models of many body systems amounts to estimate high dimensional probability distributions. It is plagued by the curse of dimensionality. Relying on a inverse renormalization group, we show that this curse is avoided by modeling energy interactions across scales. It leads to well conditioned convex interaction terms that are estimated. Multiscale models are introduced for weak lensing, mass distributions of the cosmic web and turbulence fields, with scattering covariances. They rely on covariances of wavelet coefficient modulus.
Spectacular results have been obtained by generative mdels with damped Langevion diffusions. They rely on energy gradient estimations with deep neural networks that are not well understood. The second colloquim reviews these algorithms. It relates these neural network models to multiscale renormalisation group decompositions. Generalisation and overfitting properties are explained.
There will be 2 colloquium talks and a special topic talk with the following titles:
Colloquium 1: Energy estimation and data generation by inverse renormalisaiton group
Special topic: Log-Sobolev stability, wavelets and interaction energies across scales
Colloquium 2: Multiscale neural network models for generation by score diffusion
The third and last lecture in this Series is on Friday 2 February 2024 starting at 14.00 hrs in the Budinich Lecture Hall and will be on "Multiscale neural network models for generation by score diffusion".
Light refreshments will be served after the talk. All are welcome to attend.
The lectures will also be livestreamed from the ICTP website.