Scientific Calendar Event



Description
An ICTP Hybrid meeting

This workshop will focus on the modelisation and statistical analysis of large structured data sets as appearing in modern signal processing and machine learning. Recurrent questions will be: what are "good" models of high-dimensional data which are realistic enough while remaining analytically tractable, and what are their universality properties?
 
Many modern problems (e.g. compressed sensing, community detection, PCA and tensor decomposition) seek to infer some latent signal from high-dimensional noisy data. A theoretical analysis is often challenging due to the subtle correlations and structured dependencies among the observed features. Remarkably, many of these systems exhibit universal statistics ie., similar properties as a surrogate random system. There has been substantial recent progress at the intersection of statistical physics, statistics, probability and machine learning in rigorously establishing these empirical observations, and these properties have been critically exploited for statistical learning.

This workshop will focus on these recent interdisciplinary investigations, with a view towards discovering new connections among the diverse approaches to these problems of common interest.

Topics:
  • High-dimensional statistics and inference
  • Statistical learning
  • Models of structured data
  • Universality
  • Statistical mechanics

Speakers:
S. ACEVEDO, SISSA, Italy
F. CAMILLI, ICTP, Italy
​R. DUDEJA, Harvard University, USA
Z. FAN, Yale University, USA
F. GERACE, SISSA, Italy
S. GOLDT, SISSA, Italy
A. JAGANNATH, Waterloo University, Canada
Y. KABASHIMA, Tokyo University, Japan
J. KO, ENS Lyon, France
B. LOUREIRO, ENS Paris, France
Y. LU, Harvard University, USA
M. MARSILI, ICTP, Italy
M. MONDELLI, IST, Austria
R. NICKL, Cambridge University, UK
M. OPPER, Birmingham University, UK
P. SUR, Harvard University, USA
T. TAKAHASHI, Tokyo University, Japan
S. VILLAR, Johns Hopkins University, USA


Call for Contributed Abstracts: All applicants are encouraged to submit an abstract for a poster presentation. Abstract templates are available below for download.

Grants: A limited number of grants are available to support the attendance of selected participants, with priority given to participants from developing countries. There is no registration fee.

Go to day
  • Monday, 3 July 2023
    • 08:30 - 09:30 Registration and Administrative formalities
      REGISTRATION: Upon arrival, Visitors not staying in the ICTP Guest Houses, are kindly requested to complete registration formalities at the Leonardo building (Lobby). The Registration Desk will be open from 8.30 to 9.30.
    • 09:30 - 16:10
      • 09:30 Approximately equivariant graph networks 50'
        Speaker: Soledad VILLAR (Johns Hopkins University, USA)
        Material: Abstract Video
      • 10:20 Coffee Break 30'
      • 10:50 Exploring bagging with structured data: Insights from precise asymptotics 50'
        Speaker: Takashi TAKAHASHI (Tokyo University, Japan)
        Material: Abstract Video
      • 11:40 Learning Noisy Rank-One matrices: Bayes, Non-Bayes, and Large Deviations 50'
        Speaker: Justin KO (ENS Lyon, France)
        Material: Abstract Video
      • 12:30 Lunch 1h30'
      • 14:00 From Spectral Estimators to Approximate Message Passing... And Back 50'
        Speaker: Marco MONDELLI (IST, Austria)
        Material: Abstract Video
      • 14:50 Coffee Break 30'
      • 15:20 Data that “make sense” 50'
        Speaker: Matteo MARSILI (ICTP, Italy)
        Material: Abstract Video
  • Tuesday, 4 July 2023
    • 09:30 - 16:10
      • 09:30 Gaussian Universality of Perceptrons with Random Labels 50'
        Speaker: Federica GERACE (SISSA, Italy)
        Material: Abstract Video
      • 10:20 Coffee Break 30'
      • 10:50 Universality and feature learning in two-layer neural networks 50'
        Speaker: Bruno LOUREIRO (ENS Paris, France)
        Material: Abstract Video
      • 11:40 Spectral Universality in High-Dimensional Statistics 50'
        Speaker: Rishabh DUDEJA (Harvard University, USA)
        Material: Abstract Video
      • 12:30 Lunch 1h30'
      • 14:00 Fundamental limits in structured PCA, and how to reach them 50'
        Speaker: Jean BARBIER (ICTP, Italy)
        Material: Abstract Video
      • 14:50 Coffee Break 30'
      • 15:20 Intrinsic dimension estimation in spin systems 50'
        Speaker: Santiago ACEVEDO (SISSA, Italy)
        Material: Abstract Video
  • Wednesday, 5 July 2023
    • 10:30 - 12:35
      • 10:30 Replica method with approximate inference 50'
        Speaker: Manfred OPPER (Birmingham University, UK)
        Material: Abstract Video
      • 11:20 Coffee Break 20'
      • 11:40 Compressed sensing based on diffusion models 50'
        Speaker: Yoshiyuki KABASHIMA (Tokyo University, Japan)
        Material: Abstract
      • 12:30 Group photo 5'
    • 17:00 - 20:00 Welcome reception
      A walk from Leonardo building starts at 5:00 PM & Welcome Reception starts at 6:30 PM
  • Thursday, 6 July 2023
    • 10:30 - 18:00
      • 10:30 How much does it cost to forget noise structure in low-rank matrix estimation? 50'
        Speaker: Manuel SAENZ (Universidad de La República Uruguay, Uruguay)
        Material: Abstract Video
      • 11:20 Coffee Break 20'
      • 11:40 On orthogonally-invariant spin glasses and linear models with invariant designs 50'
        Speaker: Zhou FAN (Yale University, USA)
        Material: Abstract Video
      • 12:30 Lunch 1h30'
      • 14:00 Equivalence Principles for Nonlinear Random Matrices: A Closer Look 50'
        CANCELLED
        Speaker: Yue LU (Harvard University, USA)
        Material: Abstract
      • 14:50 Coffee Break 5'
      • 14:55 The Gaussian world is not enough - how data shapes neural network representations 50'
        Speaker: Sebastian GOLDT (SISSA, Italy)
        Material: Abstract Video
      • 15:45 Poster session 1h45'
        Material: Participants abstracts
  • Friday, 7 July 2023
    • 10:30 - 12:30
      • 10:30 Debiasing regularized linear estimators with "spectrum-aware adjustments" 50'
        Speaker: Pragya SUR (Harvard University, USA)
        Material: Abstract Video
      • 11:20 Coffee Break 20'
      • 11:40 Fundamental limits of shallow neural networks with small training sets 50'
        Speaker: Francesco CAMILLI (ICTP, Italy)
        Material: Abstract Video