BEGIN:VCALENDAR
VERSION:2.0
CALSCALE:GREGORIAN
PRODID:UW-Madison-Physics-Events
BEGIN:VEVENT
SEQUENCE:2
UID:UW-Physics-Event-6925
DTSTART:20220427T160000Z
DTEND:20220427T171500Z
DTSTAMP:20221203T185137Z
LAST-MODIFIED:20220426T183415Z
LOCATION:Chamberlin 5280 (Zoom link for those attending online: https:
//uwmadison.zoom.us/j/98994425904?pwd=cnY5REVFaFY5bU1HOUN5V1ZSemdGdz09
)
SUMMARY:Quantum field theory and deep neural networks\, Physics ∩ ML
Seminar\, Ro Jefferson\, Nordita
DESCRIPTION:Recently\, exciting progress has been made in the study of
deep neural networks (DNNs) by applying ideas and techniques from phy
sics\, and in particular QFT. In this talk\, I will first give a brief
overview of some key aspects of the approach to DNNs from effective t
heory\, and highlight the information-theoretic language that unites t
hese two seemingly disparate fields. Then\, I will explain how one can
go beyond the level of analogy by explicitly constructing a bona-fide
QFT corresponding to a general class of DNNs encompassing both recurr
ent and feedforward architectures. The resulting theory closely resemb
les the well-studied O(N) vector model\, in which the variance of the
weight initializations plays the role of the 't Hooft coupling. In thi
s framework\, the Gaussian process approximation used in machine learn
ing corresponds to a free field theory\, and finite-width effects can
be computed perturbatively in the ratio of depth to width\, T/N. These
provide corrections to the correlation length that controls the depth
to which information can propagate through the network\, and thereby
sets the scale at which such networks are trainable by gradient descen
t. This analysis provides a first-principles approach to the rapidly e
merging NN-QFT correspondence\, and opens several interesting avenues
to the study of criticality in deep neural networks.
\n
\nBased on 210
9.13247 with Kevin T. Grosvenor.
URL:https://www.physics.wisc.edu/events/?id=6925
END:VEVENT
END:VCALENDAR