APSPS/統計数学セミナー 4/22

[Yuta Koike]

IBISMLの皆様

東京大学の小池です。

4月のAsia-Pacific Seminar in Probability and Statistics (APSPS)の告知です。

Date: 22 April (Wed), 2026
Time: 14:00 (Japan)
Speaker: Liam Hodgkinson (U. Melbourne)
Title: The High-Temperature Marchenko-Pastur Model for predicting deep
learning performance
Abstract: Recent theoretical successes in deep learning, such as the
celebrated neural scaling laws, are centred around the prevalence of
heavy-tailed spectral densities in Jacobians, Hessians, and weight
matrices; a phenomenon known as heavy-tailed mechanistic universality
(HT-MU). While empirical evidence suggests a robust correlation
between heavy-tailed metrics and model performance, current models
from random matrix theory (typically those designed around the
Marchenko-Pastur distribution) fail to explain these observations. In
this talk, I will propose a general family of random matrix models
derived from beta-ensembles, called the high-temperature
Marchenko-Pastur (HTMP) ensemble, to explore attributes that give rise
to heavy-tailed behaviour in trained neural networks. In HTMP,
spectral densities with power laws on (upper and lower) tails can
arise because of implicit biases in the model structure, represented
in terms of an “eigenvalue repulsion” parameter. I will show that this
parameter provides an excellent predictor of model performance and
provides the major missing piece needed to bridge the theory-practice
gap for random matrix theory in deep learning.

https://sites.google.com/view/apsps/home


東京大学からは、統計数学セミナーを通して、Zoomで配信します。
参加希望の方は以下のGoogle Formより２日前までにご登録ください。
ご登録後、会議参加に必要なURLを送付いたします。
https://docs.google.com/forms/d/e/1FAIpQLSelob9GGjU93hg6q8yXHxN2MjqUtxFkP0qprylMkqhXHRcahg/viewform

統計数学セミナーオーガナイザー
吉田 朋広(東大数理)，増田 弘毅 (東大数理)，荻原 哲平(東大情報理工), 小池 祐太(東大数理)
Nakahiro Yoshida (University of Tokyo), Hiroki Masuda (University of
Tokyo), Teppei Ogihara (University of
Tokyo), Yuta Koike (University of Tokyo)
Email:  ogi...@mist.i.u-tokyo.ac.jp
Web:  http://www.ms.u-tokyo.ac.jp/seminar/probstat/future.html





-------------------------------------------------------------------------------------


Asia-Pacific Seminar in Probability and Statistics


The seminar is created as a permanent forum for good research in the
field. It is our way to counterbalance the restrictions imposed on us
by the pandemic and to still promote scientific exchange and
cooperation. It will run using Zoom.

Topics include probabilistic models for natural phenomena, stochastic
processes and statistical inference, statistical problems in
high-dimensional spaces, asymptotic methods, statistical theory of
diversity. At the same time, there are no restrictions on other topics
of the seminar presentations.

The emphasis is on novelty, but even more so, on beauty, and clarity.
Where possible, heuristic arguments are preferred to presentation of
technical details.  Speakers are encouraged to discuss meaningful
applications and open problems.  Good reviews are welcome.
Presentations should be accessible to good postgraduate students in
probability and mathematical statistics.

The Board of the Seminar consists of

Anirvan Chakraborty (IISER, Kolkata)
Sanjay Chaudhuri (Univ. Nebraska-Lincoln)
Feng Chen (UNSW, Sydney)
Jie Yen Fan (Monash U, Melbourne)
Jesse Goodman (U Auckland)
Martin Hazelton (U Otago, Dunedin)
Mark Holmes (U Melbourne, currently at UBC, Vancouver)
Ajay Jasra (Chinese U of Hong-Kong, Shenzhen), Chair
Estate Khmaladze (VUW, Wellington)
Chenxu Li (Peking U)
Hiroki Masuda (U Tokyo)
Stephen Muirhead (U Melbourne)
Soumendu Sundar Mukherjee (ISI, Kolkata)
Teppei Ogihara (U Tokyo), Deputy Chair
Spiro Penev (UNSW, Sydney)
Rahul Roy (ISI, Delhi)
Lijiang Yang (Tsinghua U, Beijing)
Nakahiro Yoshida (U Tokyo)



The Board is responsible for the programme. All of us will be happy to
discuss details with potential speakers. Please, e-mail us if you
would like to contribute -- addresses readily obtainable on the web.

We plan to run seminars on Wednesdays, once a month. Time will vary in
between UTC+10 - UTC+12 (India) and UTC+17 - UTC+20 (New Zealand).
Specific dates will be given on the web-page
                                https://sites.google.com/view/apsps/home

-------------------------------------------------------------------------------------