BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//TYPO3/NONSGML News system (news)//EN
CALSCALE:GREGORIAN
X-WR-CALNAME;VALUE=TEXT:Past Talks
BEGIN:VTIMEZONE
TZID:Europe/Zurich
TZURL:http://tzurl.org/zoneinfo/Europe/Zurich
X-LIC-LOCATION:Europe/Zurich
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19810329T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19961027T030000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
UID:news-989@dmi.unibas.ch
DTSTAMP:20191208T130532
DTSTART;TZID=Europe/Zurich:20191218T110000
DTEND;TZID=Europe/Zurich:20191218T110000
SUMMARY:Seminar in probability theory: Nicolas Macris (EPFL)
LOCATION:Spiegelgasse 5, Room 00.003
DESCRIPTION:High-dimensional generalized linear models are basic building b
locks of current data analysis tools including multilayers neur
al networks. They arise in signal processing, statistical infer
ence, machine learning, communication theory, and other fields.
I will explain how to establish rigorously the intrinsic infor
mation-theoretic limitations of inference and learning for a cl
ass of randomly generated instances of generalized linear model
s, thus closing several old conjectures. Examples will be shown
where one can delimit regions of parameters for which the opti
mal error rates are efficiently achievable with currently known
algorithms. I will discuss how the proof technique, based on t
he recently developed adaptive interpolation method, is able to
deal with the output nonlinearity and also to some extent with
non-separable input distributions.https://probability.dmi.unib
as.ch/seminar.html
END:VEVENT
BEGIN:VEVENT
UID:news-1021@dmi.unibas.ch
DTSTAMP:20200216T193239
DTSTART;TZID=Europe/Zurich:20191218T110000
DTEND;TZID=Europe/Zurich:20191218T110000
SUMMARY:Seminar in probability theory: Debapratim Banerjee (ISI, Kolkata)
LOCATION:Spiegelgasse 1 Room 05.002
DESCRIPTION:We consider a spin system containing pure two spin Sherrington-
Kirkpatrick Hamiltonian with Curie-Weiss interaction. The model
where the spins are spherically symmetric was considered by Ba
ik and Lee and Baik et al. which shows a two dimensional phase
transition with respect to temperature and the coupling constan
t. In this paper we prove a result analogous to Baik and Lee in
the “paramagnetic regime” when the spins are i.i.d. Radema
cher. We prove the free energy in this case is asymptotically G
aussian and can be approximated by a suitable linear spectral s
tatistics. Unlike the spherical symmetric case the free energy
here can not be written as a function of the eigenvalues of the
corresponding interaction matrix. The method in this paper rel
ies on a dense sub-graph conditioning technique introduced by B
anerjee . The proof of the approximation by the linear spectral
statistics part is close to Banerjee and Ma.https://probabilit
y.dmi.unibas.ch/seminar.html [https://probability.dmi.unibas.ch
/seminar.html]
END:VEVENT
BEGIN:VEVENT
UID:news-988@dmi.unibas.ch
DTSTAMP:20191209T133116
DTSTART;TZID=Europe/Zurich:20191211T110000
DTEND;TZID=Europe/Zurich:20191211T110000
SUMMARY:Seminar in probability theory: Daniele Tantari (Scuola Normale Superiore)
LOCATION:Spiegelgasse 5, Room 00.003
DESCRIPTION:Mean-field methods fail to reconstruct the parameters of the mo
del when the dataset is clusterized. This situation is found at
low temperatures because of the emergence of multiple thermody
namic states. The paradigmatic Hopfield model is considered in
a teacher-student scenario as a problem of unsupervised learnin
g with Restricted Boltzmann Machines (RBM). For different choic
es of the priors on units and weights, the replica symmetric ph
ase diagram of random RBM’s is analyzed and in particular the
paramagnetic phase boundary is presented as directly related t
o the optimal size of the training set necessary for a good gen
eralization. The connection between the direct and inverse prob
lem is pointed out by showing that inference can be efficiently
performed by suitably adapting both standard learning techniqu
es and standard approaches to the direct problem. \r\nhttps://
probability.dmi.unibas.ch/seminar.html [https://probability.dmi
.unibas.ch/seminar.html]
END:VEVENT
BEGIN:VEVENT
UID:news-982@dmi.unibas.ch
DTSTAMP:20191201T140857
DTSTART;TZID=Europe/Zurich:20191106T110000
DTEND;TZID=Europe/Zurich:20191106T110000
SUMMARY:Seminar in probability theory: Razvan Gurau (CNRS)
LOCATION:Spiegelgasse 5, Room 00.003
DESCRIPTION:I will give an introduction to random tensors and their applica
tions. In particular I will describe an universality result for
invariant probability measures for tensors: under generic scal
ing assumptions, for large tensor sizes any invariant tensor me
asure approaches a Gaussian. I will then discuss the implicatio
ns of this result, as well as ways to avoid it.https://probabil
ity.dmi.unibas.ch/seminar.html
END:VEVENT
BEGIN:VEVENT
UID:news-954@dmi.unibas.ch
DTSTAMP:20191028T101259
DTSTART;TZID=Europe/Zurich:20191030T110000
DTEND;TZID=Europe/Zurich:20191030T110000
SUMMARY:Seminar in probability theory: Alain-Sol Sznitman (ETH Zurich)
LOCATION:Spiegelgasse 5, Room 00.003
DESCRIPTION:In this talk we will discuss some recent large deviation asympt
otics concerning the local behavior of random interlacements on
Z^d, d≥3. In particular, we will describe the link with prev
ious results concerning macroscopic holes left inside a large b
ox, by the the adequately thickened connected component of the
boundary of the box in the vacant sets of random interlacements
.https://probability.dmi.unibas.ch/seminar.html
END:VEVENT
BEGIN:VEVENT
UID:news-894@dmi.unibas.ch
DTSTAMP:20190605T113341
DTSTART;TZID=Europe/Zurich:20190617T110000
DTEND;TZID=Europe/Zurich:20190617T110000
SUMMARY:Seminar in probability theory: Shuta Nakajima (University of Nagoya)
LOCATION:Spiegelgasse 5, Room 05.001
DESCRIPTION:In this talk, we consider the discrete directed polymermodel wi
th i.i.d. environment and we study the fluctuations of theparti
tion function. It was proven by Comets and Liu that forsufficie
ntly high temperature, the fluctuations converge indistribution
towards the product of the limiting partition functionand an i
ndependent Gaussian random variable. We extend the result tothe
whole L^2-region, which is predicted to be the maximalhigh-tem
perature region where the Gaussian fluctuations should occurund
er the considered scaling. This is joint work with Clément Cos
co.
END:VEVENT
BEGIN:VEVENT
UID:news-874@dmi.unibas.ch
DTSTAMP:20190513T132303
DTSTART;TZID=Europe/Zurich:20190515T110000
DTEND;TZID=Europe/Zurich:20190515T110000
SUMMARY:Seminar in probability theory: Augusto Teixeira (IMPA)
LOCATION:Spiegelgasse 5, Room 05.002
DESCRIPTION:In this talk we will study the asymptotic behavior of a random
walk that evolves on top of a simple symmetric exclusion proces
s. This nice example of a random walk on a dynamical random env
ironment presents its own challenges due to the slow mixing pro
perties of the underlying medium. We will discuss a law of larg
e numbers that has been proved recently for this random walk. I
nterestingly, we can only prove this law of large numbers for a
ll but two exceptional densities of the exclusion process. The
main technique that we have employed is a multi-scale renormali
zation that has been derived from works in percolation theory.
END:VEVENT
BEGIN:VEVENT
UID:news-873@dmi.unibas.ch
DTSTAMP:20190503T143804
DTSTART;TZID=Europe/Zurich:20190508T110000
DTEND;TZID=Europe/Zurich:20190508T110000
SUMMARY:Seminar in probability theory: Roland Bauerschmidt (University of Cambridge)
LOCATION:
DESCRIPTION:The classical random walk isomorphism theorems relate the local
time of a random walk to the square of a Gaussian free field.
I will present non-Gaussian versions of these theorems, relatin
g hyperbolic and hemispherical sigma models (and their supersym
metric versions) to non-Markovian random walks interacting thro
ugh their local time. Applications include a short proof of the
Sabot-Tarres limiting formula for the vertex-reinforced jump p
rocess (VRJP) and a Mermin-Wagner theorem for hyperbolic sigma
models and the VRJP. This is joint work with Tyler Helmuth and
Andrew Swan.
END:VEVENT
BEGIN:VEVENT
UID:news-843@dmi.unibas.ch
DTSTAMP:20190319T145022
DTSTART;TZID=Europe/Zurich:20190415T130000
DTEND;TZID=Europe/Zurich:20190415T130000
SUMMARY:Seminar in probability theory: Marilou Gabrié (ENS)
LOCATION:
DESCRIPTION:The complexity of deep neural networks remains an obstacle to t
he understanding of their great efficiency. Their generalisatio
n ability, a priori counter intuitive, is not yet fully account
ed for. Recently an information theoretic approach was proposed
to investigate this question.Relying on the heuristic replica
method from statistical physics we present an estimator for ent
ropies and mutual informations in models of deep model networks
. Using this new tool, we test numerically the relation between
generalisation and information.
END:VEVENT
BEGIN:VEVENT
UID:news-842@dmi.unibas.ch
DTSTAMP:20190319T144848
DTSTART;TZID=Europe/Zurich:20190410T110000
DTEND;TZID=Europe/Zurich:20190410T110000
SUMMARY:Seminar in probability theory: Lenaïc Chizat (Université Paris-Sud)
LOCATION:
DESCRIPTION:The current successes achieved by neural networks are mostly dr
iven by experimental exploration of various architectures, pipe
lines, and hyper-parameters, motivated by intuition rather than
precise theories. Focusing on the optimization/training aspect
, we will see in this talk why pushing theory forward is challe
nging, but also why it matters and key insights it may lead to.
We will review some recent results on the phenomenon of "lazy
training", on the role of over-parameterization, and on trainin
g neural networks with a single hidden layer.
END:VEVENT
BEGIN:VEVENT
UID:news-841@dmi.unibas.ch
DTSTAMP:20190319T144730
DTSTART;TZID=Europe/Zurich:20190403T110000
DTEND;TZID=Europe/Zurich:20190403T110000
SUMMARY:Seminar in probability theory: Arthur Jacot (EPFL)
LOCATION:
DESCRIPTION:We show that the behaviour of a Deep Neural Network (DNN) durin
g gradient descent is described by a new kernel: the Neural Tan
gent Kernel (NTK). More precisely, as the parameters are traine
d using gradient descent, the network function (which maps the
network inputs to the network outputs) follows a so-called kern
el gradient descent w.r.t. the NTK. We prove that as the networ
k layers get wider and wider, the NTK converges to a determinis
tic limit at initialization, which stays constant during traini
ng. This implies in particular that if the NTK is positive defi
nite, the network function converges to a global minimum. The N
TK also describes how DNNs generalise outside the training set:
for a least squares cost, the network function converges in ex
pectation to the NTK kernel ridgeless regression, explaining ho
w DNNs generalise in the so-called overparametrized regime, whi
ch is at the heart of most recent developments in deep learning
.
END:VEVENT
BEGIN:VEVENT
UID:news-840@dmi.unibas.ch
DTSTAMP:20190319T144621
DTSTART;TZID=Europe/Zurich:20190327T110000
DTEND;TZID=Europe/Zurich:20190327T110000
SUMMARY:Seminar in probability theory: Levent Sagun (EPFL)
LOCATION:
DESCRIPTION:
END:VEVENT
BEGIN:VEVENT
UID:news-839@dmi.unibas.ch
DTSTAMP:20190319T144242
DTSTART;TZID=Europe/Zurich:20190320T110000
DTEND;TZID=Europe/Zurich:20190320T110000
SUMMARY:Seminar in probability theory: David Belius (Universität Basel)
LOCATION:
DESCRIPTION:This is the first talk in a five part series of talks on deep l
earning from a theoretical point of view, held jointly between
the probability theory and machine learning groups of the Depar
tment of Mathematics and Computer Science. The four invited spe
akers that follow after this talk are young researchers who are
contributing in different ways to what will hopefully eventual
ly be a comprehensive theory of deep neural networks.In this fi
rst talk I will introduce the main theoretical questions about
deep neural networks:1. Representation - what can deep neural
networks represent?2. Optimization - why and under what circum
stances can we successfully train neural networks?3. Generaliza
tion - why do deep neural networks often generalize well, desp
ite huge capacity?As a preface I will review the basic models a
nd algorithms (Neural Networks, (stochastic) gradient descent,
...) and some important concepts from machine learning (capacit
y, overfitting/underfitting, generalization, ...).
END:VEVENT
BEGIN:VEVENT
UID:news-355@dmi.unibas.ch
DTSTAMP:20181030T094303
DTSTART;TZID=Europe/Zurich:20181212T110000
DTEND;TZID=Europe/Zurich:20181212T110000
SUMMARY:Seminar in probability theory: Ioan Manolescu (Fribourg)
LOCATION:SR 000.03
DESCRIPTION:TBA
END:VEVENT
BEGIN:VEVENT
UID:news-354@dmi.unibas.ch
DTSTAMP:20181030T094136
DTSTART;TZID=Europe/Zurich:20181128T110000
DTEND;TZID=Europe/Zurich:20181128T110000
SUMMARY:Seminar in probability theory: Gaultier Lambert (University of Zurich)
LOCATION:SR 000.03
DESCRIPTION:TBA
END:VEVENT
BEGIN:VEVENT
UID:news-326@dmi.unibas.ch
DTSTAMP:20181103T093011
DTSTART;TZID=Europe/Zurich:20181121T110000
DTEND;TZID=Europe/Zurich:20181121T110000
SUMMARY:Seminar in probability theory: Antti Knowles (Geneva)
LOCATION:SR 00.003
DESCRIPTION:We consider the adjacency matrix of the Erdos-Renyi graph G(N,p
) in the supercritical regime pN > C log N for some universal
constant C. We show that the eigenvalue density is with high p
robability well approximated by the semicircle law on all spec
tral scales larger than the typical eigenvalue spacing. We als
o show that all eigenvectors are completely delocalized with h
igh probability. Both results are optimal in the sense that th
ey are known to be false for pN < log N. A key ingredient of t
he proof is a new family of large deviation estimates for mult
ilinear forms of sparse vectors. \r\n\r\nJoint work with Yukun
He and Matteo Marcozzi.
END:VEVENT
BEGIN:VEVENT
UID:news-353@dmi.unibas.ch
DTSTAMP:20181103T092837
DTSTART;TZID=Europe/Zurich:20181114T110000
DTEND;TZID=Europe/Zurich:20181114T110000
SUMMARY:Seminar in probability theory: Marius Schmidt (Basel)
LOCATION:SR 000.03
DESCRIPTION:Consider the hypercube as a graph with vertex set {0,1}^N and e
dges between two vertices if they are only one coordinate flip
apart. Choosing independent standard exponentially distribute
d lengths for all edges and asking how long the shortest direc
ted paths from (0,..,0) to (1,..,1) is defines oriented first
passage percolation on the hypercube. We will discuss the conc
eptual steps needed to answer this question to the precision o
f extremal process following the two paper series "Oriented fi
rst passage percolation in the mean field limit" by Nicola Kis
tler, Adrien Schertzer and Marius A. Schmidt: arXiv:1804.03117
and arXiv:1808.04598.
END:VEVENT
BEGIN:VEVENT
UID:news-352@dmi.unibas.ch
DTSTAMP:20181103T092627
DTSTART;TZID=Europe/Zurich:20181107T110000
DTEND;TZID=Europe/Zurich:20181107T110000
SUMMARY:Seminar in probability theory: Dominik Schröder (IST Austria)
LOCATION:SR 000.03
DESCRIPTION:For Wigner-type matrices, i.e. Hermitian random matrices with
independent, not necessarily identically distributed entries ab
ove the diagonal, we show that at any cusp singularity of the
limiting eigenvalue distribution the local eigenvalue statisti
cs are universal and form a Pearcey process. Since the density
of states typically exhibits only square root or cubic root c
usp singularities, our work complements previous results on th
e bulk and edge universality and it thus completes the resolut
ion of the Wigner-Dyson-Mehta universality conjecture for the
last remaining universality type.
END:VEVENT
BEGIN:VEVENT
UID:news-327@dmi.unibas.ch
DTSTAMP:20181030T093900
DTSTART;TZID=Europe/Zurich:20181031T110000
DTEND;TZID=Europe/Zurich:20181031T110000
SUMMARY:Seminar in probability theory: Anton Klimovsky (Duisburg-Essen)
LOCATION:SR 000.03
DESCRIPTION:Finding the (space-height) distribution of the (local) extrema
of high-dimensional strongly correlated random fields is a noto
rious hard problem with many applications. Following Fyodorov &
Sommers (2007), we focus on the Gaussian fields with isotropic
increments and take the viewpoint of statistical physics. By e
xploiting various probabilistic symmetries, we rigorously deriv
e the Fyodorov-Sommers formula for the log-partition function i
n the high-dimensional limit. The formula suggests a rich pictu
re for the distribution of the local extrema akin to the celebr
ated spherical Sherrington-Kirkpatrick model with mixed p-spin
interactions.
END:VEVENT
BEGIN:VEVENT
UID:news-325@dmi.unibas.ch
DTSTAMP:20180925T212513
DTSTART;TZID=Europe/Zurich:20180906T110000
DTEND;TZID=Europe/Zurich:20180906T120000
SUMMARY:Seminar in probability theory: Lisa Hartung (New York University)
LOCATION:SR 00.003
DESCRIPTION:It was proven by Rider and Virag that the logarithm of the char
acteristic polynomial of the Ginibre ensemble converges to a lo
garithmically correlated random field. In this talk we will see
how this connection can be established on the level if powers
of the characteristic polynomial by proving convergence to Gaus
sian multiplicative chaos. We consider the range of powers in t
he L^2 phase. \r\n\r\n(Joint work in progress with Paul Bourgad
e and Guillaume Dubach).
END:VEVENT
BEGIN:VEVENT
UID:news-306@dmi.unibas.ch
DTSTAMP:20180927T122021
DTSTART;TZID=Europe/Zurich:20180822T110000
DTEND;TZID=Europe/Zurich:20180822T120000
SUMMARY:Seminar in probability theory: Alexander Drewitz (Köln)
LOCATION:SR 00.003
DESCRIPTION:We consider two fundamental percolation models with long-range
correlations: The Gaussian free field and (the vacant set) of R
andom Interlacements. Both models have been the subject of inte
nsive research during the last years and decades, on Zd as well
as on some more general graphs. We investigate some structural
percolative properties around their critical parameters, in pa
rticular the ubiquity of the infinite components of complementa
ry phases. \r\n\r\nThis talk is based on joint works with A. Pr
évost (Köln) and P.-F. Rodriguez (Bures-sur-Yvette).
END:VEVENT
END:VCALENDAR