People by Discipline

Name	Discipline	Category	Phone	Office
Istrati Daniel	Atomic, Molecular and Optical Physics	Graduate Students	02 6585 613	Marks 100
Zigler Arie	Atomic, Molecular and Optical Physics	Emeriti	02 6585 157	Marks 103
Eisenberg Hagai	Atomic, Molecular and Optical Physics	Faculty	02 6585 228	Marks
Retzker Alex	Atomic, Molecular and Optical Physics	Faculty	02 6586 329	Danciger B 203
Svetitsky Elisha	Atomic, Molecular and Optical Physics	Graduate Students	02 6586 745	Kaplun Bldg.
Botton-Dascal Mordechai	Atomic, Molecular and Optical Physics	Faculty	02 6584 255	Popick 123
Kostinski Sarah	Atomic, Molecular and Optical Physics	Post-Docs	02 6584 648	Danciger B 208
Harpaz Yotam	Atomic, Molecular and Optical Physics	Graduate Students	02 6586 083	Danciger B 120
Kirsh Naftali	Atomic, Molecular and Optical Physics	Graduate Students	02 6586 745	Kaplun Bldg.
Ren Jie	High Energy Physics	Post-Docs	02 6584 890	Kaplun 212

Upcoming Events

Yesterday

Nonlinear Physics seminar: "Statistical mechanics of wide but finite artificial neural networks"
- Start time: 12:00pm
- End date: Wednesday, December 6th
- End time: 01:30pm
- Lecturer: Mr. Gadi Naveh (HUJI)Abstract: Over the past decade, Deep neural networks (DNNs) have brought significant advancements to various fields of science and technology. Key to their success is feature learning, enabling them to extract useful representations from raw data. However, DNNs remain challenging to interpret and lack a comprehensive analytical understanding. Recent studies revealed that DNNs in the infinite width limit are equivalent to Gaussian Processes (GPs), which are well-understoodanalytically but lack feature learning capabilities due to pre-determined features.In this thesis, we aim to bridge the gap between GPs and finite DNNs, using approaches inspired by theoretical physics. Employing statistical field theory, we treat the DNN output as a random object from a statistical ensemble. We performed extensive computer simulations of DNNs trained in both stylized and real-world settings. As for analytical methods, we first employ a perturbative approach to obtain the leading correction in 1/width to the GP limit. Subsequently, we develop a self-consistent mean-field type theory, providing a GP description of finite DNNs with a data-dependent kernel that embodies feature learning properties.
- View this event in Google Calendar
Position at Kyoto
- Start time: 02:00pm
- End date: Wednesday, December 6th
- End time: 03:00pm
- View this event in Google Calendar

Search form

People by Discipline

Pages