BEGIN:VCALENDAR
VERSION:2.0
PRODID:researchseminars.org
CALSCALE:GREGORIAN
X-WR-CALNAME:researchseminars.org
BEGIN:VEVENT
SUMMARY:Cláudia Soares (Instituto Superior Técnico and ISR)
DTSTART;VALUE=DATE-TIME:20200514T163000Z
DTEND;VALUE=DATE-TIME:20200514T173000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/1
DESCRIPTION:Title: The
learning machine and beyond: a tour for the curious\nby Cláudia Soar
es (Instituto Superior Técnico and ISR) as part of Mathematics\, Physics
and Machine Learning (IST\, Lisbon)\n\n\nAbstract\nThis talk will draw a f
ew perspectives on the broad topic of Machine Learning\, with non-speciali
sts in mind. We will go through major subfields like supervised\, unsuperv
ised\, or active learning\, never forgetting the emergent reinforcement le
arning. We will cover a few different trends over recent years\, like the
mathematically inclined Support Vector Machine\, or the empirical Deep Lea
rning.\n
LOCATION:https://researchseminars.org/talk/MPML/1/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Afonso Bandeira (ETH Zurich)
DTSTART;VALUE=DATE-TIME:20200604T163000Z
DTEND;VALUE=DATE-TIME:20200604T173000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/2
DESCRIPTION:Title: Com
putation\, statistics\, and optimization of random functions\nby Afons
o Bandeira (ETH Zurich) as part of Mathematics\, Physics and Machine Learn
ing (IST\, Lisbon)\n\n\nAbstract\nWhen faced with a data analysis\, learni
ng\, or statistical inference problem\, the amount and quality of data ava
ilable fundamentally determines whether such tasks can be performed with c
ertain levels of accuracy. Indeed\, many theoretical disciplines study lim
its of such tasks by investigating whether a dataset effectively contains
the information of interest. With the growing size of datasets however\, i
t is crucial not only that the underlying statistical task is possible\, b
ut also that is doable by means of efficient algorithms. In this talk we w
ill discuss methods aiming to establish limits of when statistical tasks a
re possible with computationally efficient methods or when there is a fund
amental Statistical-to-Computational gap in which an inference task is sta
tistically possible but inherently computationally hard.\n\nThis is intima
tely related to understanding the geometry of random functions\, with conn
ections to statistical physics\, study of spin glasses\, random geometry\;
and in an important example\, algebraic invariant theory.\n
LOCATION:https://researchseminars.org/talk/MPML/2/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Marcelo Pereyra (Heriot-Watt University)
DTSTART;VALUE=DATE-TIME:20200611T163000Z
DTEND;VALUE=DATE-TIME:20200611T173000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/3
DESCRIPTION:Title: Eff
icient Bayesian computation by proximal Markov chain Monte Carlo: when Lan
gevin meets Moreau\nby Marcelo Pereyra (Heriot-Watt University) as par
t of Mathematics\, Physics and Machine Learning (IST\, Lisbon)\n\n\nAbstra
ct\nThis talk summarises some new developments in Bayesian statistical met
hodology for performing inference in high-dimensional inverse problems wit
h an underlying convex geometry. We pay particular attention to problems r
elated to imaging sciences and to new stochastic computation methods that
tightly combine proximal convex optimisation and Markov chain Monte Carlo
sampling techniques. The new computation methods are illustrated with a ra
nge of imaging experiments\, where they are used to perform uncertainty qu
antification analyses\, automatically adjust regularisation parameters\, a
nd objectively compare alternative models in the absence of ground truth.\
n
LOCATION:https://researchseminars.org/talk/MPML/3/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Csaba Szepesvári (University of Alberta and DeepMind)
DTSTART;VALUE=DATE-TIME:20200625T163000Z
DTEND;VALUE=DATE-TIME:20200625T173000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/5
DESCRIPTION:Title: Con
fident Off-Policy Evaluation and Selection through Self-Normalized Importa
nce Weighting\nby Csaba Szepesvári (University of Alberta and DeepMin
d) as part of Mathematics\, Physics and Machine Learning (IST\, Lisbon)\n\
n\nAbstract\nOff-policy evaluation is the problem of predicting the value
of a policy given some batch of data. In the language of statistics\, this
is also called counterfactual estimation. Batch policy optimization refer
s to the problem of finding a good policy\, again\, given some logged data
.\nIn this talk\, I will consider the case of contextual bandits\, give a
brief (and incomplete) review of the approaches proposed in the literature
and explain why this problem is difficult. Then\, I will describe a new a
pproach based on self-normalized importance weighting. In this approach\,
a semi-empirical Efron-Stein concentration inequality is combined with Har
ris' inequality to arrive at non-vacuous high-probability value lower boun
ds\, which can then be used in a policy selection phase. On a number of sy
nthetic and real datasets this new approach is found to be significantly s
uperior than its main competitors\, both in terms of tightness of the conf
idence intervals and the quality of the policies chosen. \n\nThe talk is b
ased on joint work with Ilja Kuzborskij\, Claire Vernade and Andras Gyorgy
.\n
LOCATION:https://researchseminars.org/talk/MPML/5/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Kyle Cranmer (NYU)
DTSTART;VALUE=DATE-TIME:20200702T163000Z
DTEND;VALUE=DATE-TIME:20200702T173000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/6
DESCRIPTION:Title: On
the Interplay between Physics and Deep Learning.\nby Kyle Cranmer (NYU
) as part of Mathematics\, Physics and Machine Learning (IST\, Lisbon)\n\n
\nAbstract\nThe interplay between physics and deep learning is typically d
ivided into two themes.\nThe first is “physics for deep learning”\, wh
ere techniques from physics are brought to bear on understanding dynamics
of learning. The second is “deep learning for physics\,” which focuses
on application of deep learning techniques to physics problems. I will pr
esent a more nuanced view of this interplay with examples of how the struc
ture of physics problems have inspired advances in deep learning and how i
t yields insights on topics such as inductive bias\, interpretability\, an
d causality.\n
LOCATION:https://researchseminars.org/talk/MPML/6/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Hilbert Johan Kappen (Donder Institute\, Radboud University Nijmeg
en\, the Netherlands)
DTSTART;VALUE=DATE-TIME:20200528T163000Z
DTEND;VALUE=DATE-TIME:20200528T173000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/7
DESCRIPTION:Title: Pat
h integral control theory\nby Hilbert Johan Kappen (Donder Institute\,
Radboud University Nijmegen\, the Netherlands) as part of Mathematics\, P
hysics and Machine Learning (IST\, Lisbon)\n\n\nAbstract\nStochastic optim
al control theory deals with the problem to compute an optimal set of acti
ons to attain some future goal. Examples are found in many contexts such a
s motor control tasks for robotics\, planning and scheduling tasks or mana
ging a financial portfolio. The computation of the optimal control is typi
cally very difficult due to the size of the state space and the stochastic
nature of the problem. Special cases for which the computation is tractab
le are linear dynamical systems with quadratic cost and deterministic cont
rol problems. For a special class of non-linear stochastic control problem
s\, the solution can be mapped onto a statistical inference problem. For t
hese so-called path integral control problems the optimal cost-to-go solut
ion of the Bellman equation is given by the minimum of a free energy. I wi
ll give a high level introduction to the underlying theory and illustrate
with some examples from robotics and other areas.\n
LOCATION:https://researchseminars.org/talk/MPML/7/
END:VEVENT
BEGIN:VEVENT
SUMMARY:André David Mendes (CERN)
DTSTART;VALUE=DATE-TIME:20200521T163000Z
DTEND;VALUE=DATE-TIME:20200521T173000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/8
DESCRIPTION:Title: How
we discovered the Higgs ahead of schedule - ML's role in unveiling the ke
ystone of elementary particle physics\nby André David Mendes (CERN) a
s part of Mathematics\, Physics and Machine Learning (IST\, Lisbon)\n\n\nA
bstract\nn 2010\, when the LHC started colliding proton pairs in earnest\,
multi-variate analyses were newfangled methods starting to make inroads i
n experimental particle physics. These methods faced widespread skepticism
as to their performance and biases\, reflecting a winter of suspicion ove
r overtrained neural networks that set in in the late 1990s. Thanks to mor
e robust techniques\, like boosted decision trees\, it became possible to
make better and more extensive use of the full information recorded in par
ticle collisions at the Tevatron and LHC colliders.\n\nThe Higgs boson dis
covery by the CMS and ATLAS collaborations in 2012 was only possible becau
se of the use of multi-variate techniques that enhanced the sensitivity by
up to the equivalent of having 50% more collision data available for anal
ysis.\n\nWe will review the use of classification and regression in the Hi
ggs to diphoton search and subsequent discovery\, a concrete example of a
decade-old ML-based analysis in high-energy particle physics. Particular e
mphasis will be placed in the modular design of the analysis and the inher
ent explainability advantages\, used to great effect in assuaging concerns
raised by hundreds of initially-skeptical colleagues in the CMS collabora
tion.\nFinally\, we'll quickly highlight some particle physics challenges
that have contributed to\, and made use of\, the last decade of graph\, ad
versarial\, and deep ML developments.\n
LOCATION:https://researchseminars.org/talk/MPML/8/
END:VEVENT
BEGIN:VEVENT
SUMMARY:João Miranda Lemos (Instituto Superior Técnico and INESC-ID)
DTSTART;VALUE=DATE-TIME:20200716T163000Z
DTEND;VALUE=DATE-TIME:20200716T173000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/9
DESCRIPTION:Title: Rei
nforcement learning and adaptive control\nby João Miranda Lemos (Inst
ituto Superior Técnico and INESC-ID) as part of Mathematics\, Physics and
Machine Learning (IST\, Lisbon)\n\n\nAbstract\nThe aim of this seminar is
to explain\, to a wide audience\, how to combine optimal control techniqu
es with reinforcement learning\, by using approximate dynamic programming\
, and artificial neural networks\, to obtain adaptive optimal controllers.
Although with roots since the end of the XX century\, this problem has be
en the subject of an increasing attention. In addition to the promising to
ols that it offers to tackle difficult nonlinear problems with major engin
eering importance (ranging from robotics to biomedical engineering and bey
hond)\, it has the charm of creating a meeting point between the control a
nd machine learning research communities.\n
LOCATION:https://researchseminars.org/talk/MPML/9/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Francisco C. Santos (Instituto Superior Técnico and INESC-ID)
DTSTART;VALUE=DATE-TIME:20200709T163000Z
DTEND;VALUE=DATE-TIME:20200709T173000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/10
DESCRIPTION:Title: Cl
imate action and cooperation dynamics under uncertainty\nby Francisco
C. Santos (Instituto Superior Técnico and INESC-ID) as part of Mathematic
s\, Physics and Machine Learning (IST\, Lisbon)\n\n\nAbstract\nWhen attemp
ting to avoid global warming\, individuals often face a social dilemma in
which\, besides securing future benefits\, it is also necessary to reduce
the chances of future losses. In this talk\, I will resort to game theory
and populations of adaptive agents to offer a theoretical analysis of this
type of dilemmas\, in which the risk of failure plays a central role in i
ndividual decisions. I will discuss both deterministic dynamics in large p
opulations\, and stochastic social learning dynamics in finite populations
. This class of models can be shown to capture some of the essential featu
res discovered in recent key experiments while allowing one to extend in n
on-trivial ways the experimental conditions to regions of practical intere
st. Moreover\, this approach leads us to identify useful parallels between
ecological and socio-economic systems\, particularly in what concerns the
evolution and self-organization of their institutions. Particularly\, our
results suggest that global coordination for a common good should be atte
mpted through a polycentric structure of multiple small-scale agreements\,
in which perception of risk is high and uncertainty in collective goals i
s minimized. Whenever the perception of risk is low\, our results indicate
that sanctioning institutions may significantly enhance the chances of co
ordinating to tame the planet's climate\, as long as they are implemented
in a bottom-up manner. I will discuss the impact on public goods dilemmas
of heterogeneous political networks and wealth inequality\, including dis
tribution of wealth representative of existing inequalities among nations.
Finally\, I will briefly discuss the impact of scientific uncertainty —
both in what concerns the collective targets and the time window availabl
e for action — on individuals' strategies and polarization of preference
s.\n
LOCATION:https://researchseminars.org/talk/MPML/10/
END:VEVENT
BEGIN:VEVENT
SUMMARY:João Xavier (Instituto Superior Técnico and ISR)
DTSTART;VALUE=DATE-TIME:20200618T163000Z
DTEND;VALUE=DATE-TIME:20200618T173000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/11
DESCRIPTION:Title: Le
arning from distributed datasets: an introduction with two examples\nb
y João Xavier (Instituto Superior Técnico and ISR) as part of Mathematic
s\, Physics and Machine Learning (IST\, Lisbon)\n\n\nAbstract\nData are in
creasingly measured\, in ever tinier minutiae\, by networks of spatially d
istributed agents. Illustrative examples include a team of robots searchin
g a large region\, a collection of sensors overseeing a critical infra-str
ucture\, or a swarm of drones policing a wide area.\n\nHow to learn from t
hese large\, spatially distributed datasets? In the centralized approach e
ach agent forwards its dataset to a fusion center\, which then carries out
the learning from the pile of amassed datasets. This approach\, however\,
prevents the number of agents to scale up: as more and more agents ship d
ata to the center\, not only the communication channels near the center qu
ickly swell to congestion\, but also the computational power of the center
is rapidly outpaced.\n\nIn this seminar\, I describe the alternative appr
oach of distributed learning. Here\, no fusion center exists\, and the age
nts themselves recreate the centralized computation by exchanging short me
ssages (not data) between network neighbors. To illustrate\, I describe tw
o learning algorithms: one solves convex learning problems via a token tha
t randomly roams through the network\, and the other solves a classificati
on problem via random meetings between agents (e.g.\, gossip)\, each agent
measuring only its own stream of features.\n\nThis seminar is aimed at no
n-specialists. Rather than trying to impart the latest developments of the
field\, I hope to open a welcoming door to those wishing to have a peek a
t this bubbling field of research\, where optimization\, control\, probabi
lity\, and machine learning mingle happily.\n
LOCATION:https://researchseminars.org/talk/MPML/11/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Marylou Gabrié (Center for Data Science\, NYU and Flatiron Instit
ute\, CCM)
DTSTART;VALUE=DATE-TIME:20200723T163000Z
DTEND;VALUE=DATE-TIME:20200723T173000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/12
DESCRIPTION:Title: Pr
ogress and hurdles in the statistical mechanics of deep learning\nby M
arylou Gabrié (Center for Data Science\, NYU and Flatiron Institute\, CCM
) as part of Mathematics\, Physics and Machine Learning (IST\, Lisbon)\n\n
\nAbstract\nUnderstanding the great performances of deep neural networks i
s a very active direction of research with contributions coming from a wid
e variety of fields. The statistical mechanics of learning is a theoretica
l framework dating back to the 80s studying learning problems from a physi
cist viewpoint and using tools from the physics of disordered systems. In
this talk\, I will first go over this traditional framework\, which relies
on the teacher-student scenario\, bayesian analysis and mean-field approx
imations. Then I will discuss some recent advances in the corresponding an
alysis of modern deep neural network\, and highlight remaining challenges.
\n
LOCATION:https://researchseminars.org/talk/MPML/12/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Masoud Mohseni (Google Quantum Artificial Intelligence Laboratory)
DTSTART;VALUE=DATE-TIME:20200730T163000Z
DTEND;VALUE=DATE-TIME:20200730T173000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/13
DESCRIPTION:Title: Te
nsorFlow Quantum: An open source framework for hybrid quantum-classical ma
chine learning.\nby Masoud Mohseni (Google Quantum Artificial Intellig
ence Laboratory) as part of Mathematics\, Physics and Machine Learning (IS
T\, Lisbon)\n\n\nAbstract\nIn this talk\, I introduce TensorFlow Quantum (
TFQ)\, an open source library that was launched by Google in March 2020\,
for the rapid prototyping of hybrid quantum-classical models for classical
or quantum data.This framework offers high-level abstractions for the des
ign\, training\, and testing of both discriminative and generative quantum
models under TensorFlow and supports high-performance quantum circuit sim
ulators. I provide an overview of the software architecture and building b
locks through several examples and illustrate TFQ functionalities via cons
tructing hybrid quantum-classical convolutional neural networks for quantu
m state classification.\n
LOCATION:https://researchseminars.org/talk/MPML/13/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Gunnar Carlsson (Stanford University)
DTSTART;VALUE=DATE-TIME:20200930T170000Z
DTEND;VALUE=DATE-TIME:20200930T180000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/14
DESCRIPTION:Title: To
pological Data Analysis and Deep Learning\nby Gunnar Carlsson (Stanfor
d University) as part of Mathematics\, Physics and Machine Learning (IST\,
Lisbon)\n\n\nAbstract\nDeep Learning is a powerful collection of techniqu
es for statistical learning\, which has shown dramatic applications in man
y different directions\, including including the study of data sets of ima
ges\, text\, and time series. It uses neural networks\, specifically convo
lutional neural networks (CNN's)\, to produce these results. What we have
observed recently is that methods of topology can contribute to this effor
t\, in diagnosing behavior within the CNN's\, in the design of neural netw
orks with excellent computational properties\, and in improving generaliza
tion\, i.e. the transfer of results of one neural network from one data se
t to another of similar type. We'll discuss topological methods in data sc
ience\, as well as there application to this interesting set of techniques
.\n
LOCATION:https://researchseminars.org/talk/MPML/14/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Lindsey Gray (Fermi National Accelerator Laboratory)
DTSTART;VALUE=DATE-TIME:20201014T170000Z
DTEND;VALUE=DATE-TIME:20201014T180000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/15
DESCRIPTION:Title: Gr
aph Neural Networks for Pattern Recognition in Particle Physics\nby Li
ndsey Gray (Fermi National Accelerator Laboratory) as part of Mathematics\
, Physics and Machine Learning (IST\, Lisbon)\n\n\nAbstract\nModern partic
le physics detectors generate copious amounts of data packed with meaning
that provides the means for high-quality measurements in demanding experim
ental environments. To achieve these measurements there is a trend towards
finer granularity in these detectors and that implies the data read out h
as less intrinsic structure. Accurate pattern recognition is required to d
efine the signatures of particles within those detectors and simultaneousl
y extract physical parameters for the particles. Typically\, algorithms to
achieve these goals are written using well known unsupervised algorithms\
, but recent advances in machine learning on graph structures\, "Graph Neu
ral Networks" (GNNs)\, provide powerful new methodologies for designing pa
ttern recognition algorithms. In particular\, methodologies for predicting
the link structure between pieces of data from detectors are well suited
to the particle physics pattern recognition task. Furthermore\, there are
interesting avenues for enforcing known symmetries of the data into the ou
tput of such networks and there is ongoing research in this direction. Thi
s talk will discuss the challenges of pattern recognition\, the advent of
GNNs and the connections to particle physics\, and the paths of research a
head for fully utilizing this powerful new tool.\n
LOCATION:https://researchseminars.org/talk/MPML/15/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Carola-Bibiane Schönlieb (DAMTP\, University of Cambridge)
DTSTART;VALUE=DATE-TIME:20201120T150000Z
DTEND;VALUE=DATE-TIME:20201120T160000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/16
DESCRIPTION:Title: Co
mbining knowledge and data driven methods for solving inverse imaging prob
lems - getting the best from both worlds\nby Carola-Bibiane Schönlieb
(DAMTP\, University of Cambridge) as part of Mathematics\, Physics and Ma
chine Learning (IST\, Lisbon)\n\n\nAbstract\nInverse problems in imaging r
ange from tomographic reconstruction (CT\, MRI\, etc) to image deconvoluti
on\, segmentation\, and classification\, just to name a few. In this talk
I will discuss\napproaches to inverse imaging problems which have both a m
athematical modelling (knowledge driven) and a machine learning (data-driv
en) component. Mathematical modelling is crucial in the presence of ill-po
sedness\, making use of information about the imaging data\, for narrowing
down the search space. Such an approach results in highly generalizable r
econstruction and analysis methods which come with desirable solutions gua
rantees. Machine learning on the other hand is a powerful tool for customi
sing methods to individual data sets. Highly parametrised models such as d
eep neural networks in particular\, are powerful tools for accurately mode
lling prior information about solutions. The combination of these two para
digms\, getting the best from both of these worlds\, is the topic of this
talk\, furnished with examples for image classification under minimal supe
rvision and for tomographic image reconstruction.\n
LOCATION:https://researchseminars.org/talk/MPML/16/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Weinan E (Princeton University)
DTSTART;VALUE=DATE-TIME:20201007T100000Z
DTEND;VALUE=DATE-TIME:20201007T110000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/17
DESCRIPTION:Title: Ma
chine Learning and Scientific Computing\nby Weinan E (Princeton Univer
sity) as part of Mathematics\, Physics and Machine Learning (IST\, Lisbon)
\n\n\nAbstract\nNeural network-based deep learning is capable of approxima
ting functions in very high dimension with unprecedented efficiency and ac
curacy. This has opened up many exciting new possibilities\, not just in t
raditional areas of artificial intelligence\, but also in scientific compu
ting and computational science. At the same time\, deep learning has also
acquired the reputation of being a set of “black box” type of tricks\,
without fundamental principles. This has been a real obstacle for making
further progress in machine learning.\n\nIn this talk\, I will try to addr
ess the following two questions:\n\n1. How machine learning will impact co
mputational mathematics and computational science?\n\n2. How computational
mathematics\, particularly numerical analysis\, can impact machine learni
ng? We describe some of the most important progresses that have been made
on these issues so far.\n\nOur hope is to put things into a perspective th
at will help to integrate machine learning with computational science.\n
LOCATION:https://researchseminars.org/talk/MPML/17/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Gitta Kutyniok (Mathematical Institute of the University of Munich
)
DTSTART;VALUE=DATE-TIME:20201202T180000Z
DTEND;VALUE=DATE-TIME:20201202T190000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/18
DESCRIPTION:Title: De
ep Learning meets Physics: Taking the Best out of Both Worlds in Imaging S
cience\nby Gitta Kutyniok (Mathematical Institute of the University of
Munich) as part of Mathematics\, Physics and Machine Learning (IST\, Lisb
on)\n\n\nAbstract\nPure model-based approaches are today often insufficien
t for solving complex inverse problems in imaging. At the same time\, we w
itness the tremendous success of data-based methodologies\, in particular\
, deep neural networks for such problems. However\, pure deep learning app
roaches often neglect known and valuable information from physics.\n\nIn t
his talk\, we will provide an introduction to this problem complex and the
n discuss a general conceptual approach to inverse problems in imaging\, w
hich combines deep learning and physics. This hybrid approach is based on
shearlet-based sparse regularization and deep learning and is guided by a
microlocal analysis viewpoint to pay particular attention to the singulari
ty structures of the data. Finally\, we will present several applications
such as tomographic reconstruction and show that our approach outperforms
previous methodologies\, including methods entirely based on deep learning
.\n
LOCATION:https://researchseminars.org/talk/MPML/18/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Tommaso Dorigo (Italian Institute for Nuclear Physics)
DTSTART;VALUE=DATE-TIME:20201125T180000Z
DTEND;VALUE=DATE-TIME:20201125T190000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/19
DESCRIPTION:Title: De
aling with Systematic Uncertainties in HEP Analysis with Machine Learning
Methods\nby Tommaso Dorigo (Italian Institute for Nuclear Physics) as
part of Mathematics\, Physics and Machine Learning (IST\, Lisbon)\n\n\nAbs
tract\nI will discuss the impact of nuisance parameters on the effectivene
ss of supervised classification in high energy physics problems\, and tech
niques that may mitigate or remove their effect in the search for optimal
selection criteria and variable transformations. The approaches discussed
include nuisance parametrized models\, modified or adversary losses\, semi
supervised learning approaches and inference-aware techniques.\n
LOCATION:https://researchseminars.org/talk/MPML/19/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Florent Krzakala (EPFL)
DTSTART;VALUE=DATE-TIME:20201028T180000Z
DTEND;VALUE=DATE-TIME:20201028T190000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/20
DESCRIPTION:Title: So
me exactly solvable models for statistical machine learning\nby Floren
t Krzakala (EPFL) as part of Mathematics\, Physics and Machine Learning (I
ST\, Lisbon)\n\n\nAbstract\nThe increasing dimensionality of data in the m
odern machine learning age presents new challenges and opportunities. The
high-dimensional settings allow one to use powerful asymptotic methods fro
m probability theory and statistical physics to obtain precise characteriz
ations and develop new algorithmic approaches. There is indeed a decades-l
ong tradition in statistical physics with building and solving such simpli
fied models of neural networks.\n\nI will give examples of recent works th
at build on powerful methods of physics of disordered systems to analyze d
ifferent problems in machine learning and neural networks\, including over
parameterization\, kernel methods\, and the gradient descent algorithm in
a high dimensional non-convex setting.\n
LOCATION:https://researchseminars.org/talk/MPML/20/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Joan Bruna (Courant Institute and Center for Data Science\, NYU)
DTSTART;VALUE=DATE-TIME:20201104T180000Z
DTEND;VALUE=DATE-TIME:20201104T190000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/21
DESCRIPTION:Title: Ma
thematical aspects of neural network learning through measure dynamics
\nby Joan Bruna (Courant Institute and Center for Data Science\, NYU) as p
art of Mathematics\, Physics and Machine Learning (IST\, Lisbon)\n\n\nAbst
ract\nHigh-dimensional learning remains an outstanding phenomena where exp
erimental evidence outpaces our current mathematical understanding\, mostl
y due to the recent empirical successes of Deep Learning algorithms. Neura
l Networks provide a rich yet intricate class of functions with statistica
l abilities to break the curse of dimensionality\, and where physical prio
rs can be tightly integrated into the architecture to improve sample effic
iency. Despite these advantages\, an outstanding theoretical challenge in
these models is computational\, ie providing an analysis that explains suc
cessful optimization and generalization in the face of existing worst-case
computational hardness results.\n\nIn this talk\, I will focus on the fra
mework that lifts parameter optimization to an appropriate measure space.
I will cover existing results that guarantee global convergence of the res
ulting Wasserstein gradient flows\, as well as recent results that study t
ypical fluctuations of the dynamics around their mean field evolution. We
will also discuss extensions of this framework beyond vanilla supervised l
earning\, to account for symmetries in the function\, as well as for compe
titive optimization.\n
LOCATION:https://researchseminars.org/talk/MPML/21/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Bin Dong (BICMR\, Peking University)
DTSTART;VALUE=DATE-TIME:20201111T110000Z
DTEND;VALUE=DATE-TIME:20201111T120000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/22
DESCRIPTION:Title: Le
arning and Learning to Solve PDEs\nby Bin Dong (BICMR\, Peking Univers
ity) as part of Mathematics\, Physics and Machine Learning (IST\, Lisbon)\
n\n\nAbstract\nDeep learning continues to dominate machine learning and ha
s been successful in computer vision\, natural language processing\, etc.
Its impact has now expanded to many research areas in science and engineer
ing. In this talk\, I will mainly focus on some recent impact of deep lear
ning on computational mathematics. I will present our recent work on bridg
ing deep neural networks with numerical differential equations. On the one
hand\, I will show how to design transparent deep convolutional networks
to uncover hidden PDE models from observed dynamical data. On the other ha
nd\, I will present our preliminary attempt to establish a deep reinforcem
ent learning based framework to solve 1D scalar conservation laws\, and a
meta-learning approach for solving linear parameterized PDEs based on the
multigrid method.\n
LOCATION:https://researchseminars.org/talk/MPML/22/
END:VEVENT
BEGIN:VEVENT
SUMMARY:René Vidal (Mathematical Institute for Data Science\, Johns Hopki
ns University)
DTSTART;VALUE=DATE-TIME:20201216T180000Z
DTEND;VALUE=DATE-TIME:20201216T190000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/23
DESCRIPTION:Title: Fr
om Optimization Algorithms to Dynamical Systems and Back\nby René Vid
al (Mathematical Institute for Data Science\, Johns Hopkins University) as
part of Mathematics\, Physics and Machine Learning (IST\, Lisbon)\n\n\nAb
stract\nRecent work has shown that tools from dynamical systems can be use
d to analyze accelerated optimization algorithms. For example\, it has bee
n shown that the continuous limit of Nesterov’s accelerated gradient (NA
G) gives an ODE whose convergence rate matches that of NAG for convex\, un
constrained\, and smooth problems. Conversely\, it has been shown that NAG
can be obtained as the discretization of an ODE\, however since different
discretizations lead to different algorithms\, the choice of the discreti
zation becomes important. The first part of this talk will extend this typ
e of analysis to convex\, constrained and non-smooth problems by using Lya
punov stability theory to analyze continuous limits of the Alternating Dir
ection Method of Multipliers (ADMM). The second part of this talk will sho
w that many existing and new optimization algorithms can be obtained by su
itably discretizing a dissipative Hamiltonian. As an example\, we will pre
sent a new method called Relativistic Gradient Descent (RGD)\, which empir
ically outperforms momentum\, RMSprop\, Adam and AdaGrad on several non-co
nvex\nproblems.\n\nThis is joint work with Guilherme Franca\, Daniel Robin
son and Jeremias Sulam.\n
LOCATION:https://researchseminars.org/talk/MPML/23/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Mauro Maggioni (Johns Hopkins University)
DTSTART;VALUE=DATE-TIME:20201021T170000Z
DTEND;VALUE=DATE-TIME:20201021T180000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/24
DESCRIPTION:Title: Le
arning Interaction laws in particle- and agent-based systems\nby Mauro
Maggioni (Johns Hopkins University) as part of Mathematics\, Physics and
Machine Learning (IST\, Lisbon)\n\n\nAbstract\nInteracting agent-based sys
tems are ubiquitous in science\, from modeling of particles in Physics to
prey-predator and colony models in Biology\, to opinion dynamics in econom
ics and social sciences. Oftentimes the laws of interactions between the a
gents are quite simple\, for example they depend only on pairwise interact
ions\, and only on pairwise distance in each interaction. We consider the
following inference problem for a system of interacting particles or agent
s: given only observed trajectories of the agents in the system\, can we l
earn what the laws of interactions are? We would like to do this without a
ssuming any particular form for the interaction laws\, i.e. they might be
"any" function of pairwise distances. We consider this problem both the me
an-field limit (i.e. the number of particles going to infinity) and in the
case of a finite number of agents\, with an increasing number of observat
ions\, albeit in this talk we will mostly focus on the latter case. We cas
t this as an inverse problem\, and study it in the case where the interact
ion is governed by an (unknown) function of pairwise distances. We discuss
when this problem is well-posed\, and we construct estimators for the int
eraction kernels with provably good statistically and computational proper
ties. We measure their performance on various examples\, that include exte
nsions to agent systems with different types of agents\, second-order syst
ems\, and families of systems with parametric interaction kernels. We also
conduct numerical experiments to test the large time behavior of these sy
stems\, especially in the cases where they exhibit emergent behavior.\n\nT
his is joint work with F. Lu\, J.Miller\, S. Tang and M. Zhong.\n
LOCATION:https://researchseminars.org/talk/MPML/24/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Xavier Bresson (Nanyang Technological University)
DTSTART;VALUE=DATE-TIME:20210127T110000Z
DTEND;VALUE=DATE-TIME:20210127T120000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/25
DESCRIPTION:Title: Be
nchmarking Graph Neural Networks\nby Xavier Bresson (Nanyang Technolog
ical University) as part of Mathematics\, Physics and Machine Learning (IS
T\, Lisbon)\n\n\nAbstract\nGraph neural networks (GNNs) have become the st
andard toolkit for analyzing and learning from data on graphs. As the fiel
d grows\, it becomes critical to identify key architectures and validate n
ew ideas that generalize to larger\, more complex datasets. Unfortunately\
, it has been increasingly difficult to gauge the effectiveness of new mod
els in the absence of a standardized benchmark with consistent experimenta
l settings. In this work\, we introduce a reproducible GNN benchmarking fr
amework\, with the facility for researchers to add new models conveniently
for arbitrary datasets. We demonstrate the usefulness of our framework by
presenting a principled investigation into the recent Weisfeiler-Lehman G
NNs (WL-GNNs) compared to message passing-based graph convolutional networ
ks (GCNs) for a variety of graph tasks with medium-scale datasets.\n
LOCATION:https://researchseminars.org/talk/MPML/25/
END:VEVENT
BEGIN:VEVENT
SUMMARY:James Halverson (Northeastern University)
DTSTART;VALUE=DATE-TIME:20210120T180000Z
DTEND;VALUE=DATE-TIME:20210120T190000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/26
DESCRIPTION:Title: Ne
ural Networks and Quantum Field Theory\nby James Halverson (Northeaste
rn University) as part of Mathematics\, Physics and Machine Learning (IST\
, Lisbon)\n\n\nAbstract\nIn this talk I will review essentials of quantum
field theory (QFT) and demonstrate how the function-space distribution of
many neural networks (NNs) shares similar properties. This allows\, for in
stance\, computation of correlators of neural network outputs in terms of
Feynman diagrams and a direct analogy between non-Gaussian corrections in
NN distributions and particle interactions. Some cases yield divergences i
n perturbation theory\, requiring the introduction of regularization and r
enormalization. Potential advantages of this perspective will be discussed
\, including a duality between function-space and parameter-space descript
ions of neural networks.\n
LOCATION:https://researchseminars.org/talk/MPML/26/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Anna C. Gilbert (Yale University)
DTSTART;VALUE=DATE-TIME:20210113T180000Z
DTEND;VALUE=DATE-TIME:20210113T190000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/27
DESCRIPTION:Title: Me
tric representations: Algorithms and Geometry\nby Anna C. Gilbert (Yal
e University) as part of Mathematics\, Physics and Machine Learning (IST\,
Lisbon)\n\n\nAbstract\nGiven a set of distances amongst points\, determin
ing what metric representation is most "consistent" with the input distanc
es or the metric that best captures the relevant geometric features of the
data is a key step in many machine learning algorithms. In this talk\, we
focus on 3 specific metric constrained problems\, a class of optimization
problems with metric constraints: metric nearness (Brickell et al. (2008)
)\, weighted correlation clustering on general graphs (Bansal et al. (2004
))\, and metric learning (Bellet et al. (2013)\; Davis et al. (2007)).\n\n
Because of the large number of constraints in these problems\, however\, t
hese and other researchers have been forced to restrict either the kinds o
f metrics learned or the size of the problem that can be solved. We provid
e an algorithm\, PROJECT AND FORGET\, that uses Bregman projections with c
utting planes\, to solve metric constrained problems with many (possibly e
xponentially) inequality constraints. We also prove that our algorithm con
verges to the global optimal solution. Additionally\, we show that the opt
imality error decays asymptotically at an exponential rate. We show that u
sing our method we can solve large problem instances of three types of met
ric constrained problems\, out-performing all state of the art methods wit
h respect to CPU times and problem sizes.\n\nFinally\, we discuss the adap
tation of PROJECT AND FORGET to specific types of metric constraints\, nam
ely tree and hyperbolic metrics.\n
LOCATION:https://researchseminars.org/talk/MPML/27/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Caroline Uhler (MIT and Institute for Data\, Systems and Society)
DTSTART;VALUE=DATE-TIME:20210210T180000Z
DTEND;VALUE=DATE-TIME:20210210T190000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/28
DESCRIPTION:Title: Ca
usal Inference and Overparameterized Autoencoders in the Light of Drug Rep
urposing for SARS-CoV-2\nby Caroline Uhler (MIT and Institute for Data
\, Systems and Society) as part of Mathematics\, Physics and Machine Learn
ing (IST\, Lisbon)\n\n\nAbstract\nMassive data collection holds the promis
e of a better understanding of complex phenomena and ultimately\, of bette
r decisions. An exciting opportunity in this regard stems from the growing
availability of perturbation / intervention data (drugs\, knockouts\, ove
rexpression\,\netc.) in biology. In order to obtain mechanistic insights f
rom such data\, a major challenge is the development of a framework that i
ntegrates observational and interventional data and allows predicting the
effect of yet unseen interventions or transporting the effect of intervent
ions observed in one context to another. I will present a framework for ca
usal structure discovery based on such data and highlight the role of over
parameterized autoencoders. We end by demonstrating how these ideas can be
applied for drug repurposing in the current SARS-CoV-2 crisis.\n
LOCATION:https://researchseminars.org/talk/MPML/28/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Samantha Kleinberg (Stevens Institute of Technology)
DTSTART;VALUE=DATE-TIME:20201209T180000Z
DTEND;VALUE=DATE-TIME:20201209T190000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/29
DESCRIPTION:Title: Da
ta\, Decisions\, and You: Making Causality Useful and Usable in a Complex
World\nby Samantha Kleinberg (Stevens Institute of Technology) as part
of Mathematics\, Physics and Machine Learning (IST\, Lisbon)\n\n\nAbstrac
t\nThe collection of massive observational datasets has led to unprecedent
ed opportunities for causal inference\, such as using electronic health re
cords to identify risk factors for disease. However\, our ability to under
stand these complex data sets has not grown the same pace as our ability t
o collect them. While causal inference has traditionally focused on pairwi
se relationships between variables\, biological systems are highly complex
and knowing when events may happen is often as important as knowing wheth
er they will. In the first half of this talk I discuss new methods that al
low causal relationships to be reliably inferred from complex observationa
l data\, motivated by analysis of intensive care unit and other medical da
ta. Causes are useful because they allow us to take action\, but how there
is a gap between the output of machine learning and what helps people mak
e decisions. In the second part of this talk I discuss our recent findings
in testing just how people fare when using the output of machine learning
and how we can go from data to knowledge to decisions.\n
LOCATION:https://researchseminars.org/talk/MPML/29/
END:VEVENT
BEGIN:VEVENT
SUMMARY:A. Pedro Aguiar (Faculdade de Engenharia\, Universidade do Porto)
DTSTART;VALUE=DATE-TIME:20210303T180000Z
DTEND;VALUE=DATE-TIME:20210303T190000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/31
DESCRIPTION:Title: Mo
del based control design combining Lyapunov and optimization tools: Exampl
es in the area of motion control of autonomous robotic vehicles\nby A.
Pedro Aguiar (Faculdade de Engenharia\, Universidade do Porto) as part of
Mathematics\, Physics and Machine Learning (IST\, Lisbon)\n\n\nAbstract\n
The past few decades have witnessed a significant research effort in the f
ield of Lyapunov model based control design. In parallel\, optimal control
and optimization model based design have also expanded their range of app
lications\, and nowadays\, receding horizon approaches can be considered a
mature field for particular classes of control systems.\nIn this talk\, I
will argue that Lyapunov based techniques play an important role for anal
ysis of model based optimization methodologies and moreover\, both approac
hes can be combined for control design resulting in powerful frameworks wi
th formal guarantees of robustness\, stability\, performance\, and safety.
Illustrative examples in the area of motion control of autonomous robotic
vehicles will be presented for Autonomous Underwater Vehicles (AUVs)\, Au
tonomous Surface Vehicles (ASVs) and Unmanned Aerial Vehicles (UAVs).\n
LOCATION:https://researchseminars.org/talk/MPML/31/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Sanjeev Arora (Computer Science Department\, Princeton University)
DTSTART;VALUE=DATE-TIME:20210106T180000Z
DTEND;VALUE=DATE-TIME:20210106T190000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/32
DESCRIPTION:Title: Th
e quest for mathematical understanding of deep learning\nby Sanjeev Ar
ora (Computer Science Department\, Princeton University) as part of Mathem
atics\, Physics and Machine Learning (IST\, Lisbon)\n\n\nAbstract\nDeep le
arning has transformed Machine Learning and Artificial Intelligence in the
past decade. It raises fundamental questions for mathematics and theory o
f computer science\, since it relies upon solving large-scale nonconvex pr
oblems via gradient descent and its variants. This talk will be an introdu
ction to mathematical questions raised by deep learning\, and some partial
understanding obtained in recent years with respect to optimization\, gen
eralization\, self-supervised learning\, privacy etc.\n
LOCATION:https://researchseminars.org/talk/MPML/32/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Jan Peters (Technische Universitaet Darmstadt)
DTSTART;VALUE=DATE-TIME:20210423T130000Z
DTEND;VALUE=DATE-TIME:20210423T140000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/33
DESCRIPTION:Title: Ro
bot Learning - Quo Vadis?\nby Jan Peters (Technische Universitaet Darm
stadt) as part of Mathematics\, Physics and Machine Learning (IST\, Lisbon
)\n\nInteractive livestream: https://videoconf-colibri.zoom.us/j/915997596
79\nPassword hint: Register via URL in the comments at least 2h before\n\n
Abstract\nAutonomous robots that can assist humans in situations of daily
life have been a long standing vision of robotics\, artificial intelligenc
e\, and cognitive sciences. A first step towards this goal is to create ro
bots that can learn tasks triggered by environmental context or higher lev
el instruction. However\, learning techniques have yet to live up to this
promise as only few methods manage to scale to high-dimensional manipulato
r or humanoid robots. In this talk\, we investigate a general framework su
itable for learning motor skills in robotics which is based on the princip
les behind many analytical robotics approaches. It involves generating a r
epresentation of motor skills by parameterized motor primitive policies ac
ting as building blocks of movement generation\, and a learned task module
that transforms these movements into motor commands. We discuss learning
on three different levels of abstraction\, i.e.\, learning for accurate co
ntrol is needed to execute\, learning of motor primitives is needed to acq
uire simple movements\, and learning of the task-dependent „hyperparamet
ers“ of these motor primitives allows learning complex tasks. We discuss
task-appropriate learning approaches for imitation learning\, model learn
ing and reinforcement learning for robots with many degrees of freedom. Em
pirical evaluations on a several robot systems illustrate the effectivenes
s and applicability to learning control on an anthropomorphic robot arm. T
hese robot motor skills range from toy examples (e.g.\, paddling a ball\,
ball-in-a-cup\, juggling) to playing robot table tennis against a human be
ing and manipulation of various objects.\n
LOCATION:https://researchseminars.org/talk/MPML/33/
URL:https://videoconf-colibri.zoom.us/j/91599759679
END:VEVENT
BEGIN:VEVENT
SUMMARY:Miguel Couceiro (Université de Lorraine)
DTSTART;VALUE=DATE-TIME:20210203T180000Z
DTEND;VALUE=DATE-TIME:20210203T190000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/34
DESCRIPTION:Title: Ma
king ML Models fairer through explanations\, feature dropout\, and aggrega
tion\nby Miguel Couceiro (Université de Lorraine) as part of Mathemat
ics\, Physics and Machine Learning (IST\, Lisbon)\n\n\nAbstract\nAlgorithm
ic decisions are now being used on a daily basis\, and based on Machine Le
arning (ML) processes that may be complex and biased. This raises several
concerns given the critical impact that biased decisions may have on indiv
iduals or on society as a whole. Not\nonly unfair outcomes affect human ri
ghts\, they also undermine public trust in ML and AI. In this talk\, we wi
ll address fairness issues of ML models based on decision outcomes\, and w
e will show how the simple idea of "feature dropout" followed by an "ensem
ble approach" can improve model fairness without compromising its accuracy
. To illustrate we will present a general workflow that relies on explaine
rs to tackle "process fairness"\, which essentially measures a model's rel
iance on sensitive or discriminatory features. We will present different a
pplications and empirical settings that show improvements not only with re
spect to process fairness but also other fairness metrics.\n
LOCATION:https://researchseminars.org/talk/MPML/34/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Steve Brunton (University of Washington)
DTSTART;VALUE=DATE-TIME:20210331T170000Z
DTEND;VALUE=DATE-TIME:20210331T180000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/35
DESCRIPTION:Title: Ma
chine learning for Fluid Mechanics\nby Steve Brunton (University of Wa
shington) as part of Mathematics\, Physics and Machine Learning (IST\, Lis
bon)\n\n\nAbstract\nMany tasks in fluid mechanics\, such as design optimiz
ation and control\, are challenging because fluids are nonlinear and exhib
it a large range of scales in both space and time. This range of scales ne
cessitates exceedingly high-dimensional measurements and computational dis
cretization to resolve all relevant features\, resulting in vast data sets
and time-intensive computations. Indeed\, fluid dynamics is one of the or
iginal big data fields\, and many high-performance computing architectures
\, experimental measurement techniques\, and advanced data processing and
visualization algorithms were driven by decades of research in fluid mecha
nics. Machine learning constitutes a growing set of powerful techniques to
extract patterns and build models from this data\, complementing the exis
ting theoretical\, numerical\, and experimental efforts in fluid mechanics
. In this talk\, we will explore current goals and opportunities for machi
ne learning in fluid mechanics\, and we will highlight a number of recent
technical advances. Because fluid dynamics is central to transportation\,
health\, and defense systems\, we will emphasize the importance of machine
learning solutions that are interpretable\, explainable\, generalizable\,
and that respect known physics.\n
LOCATION:https://researchseminars.org/talk/MPML/35/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Hsin Yuan Huang\, (Robert) (Caltech)
DTSTART;VALUE=DATE-TIME:20210317T180000Z
DTEND;VALUE=DATE-TIME:20210317T190000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/36
DESCRIPTION:Title: In
formation-theoretic bounds on quantum advantage in machine learning\nb
y Hsin Yuan Huang\, (Robert) (Caltech) as part of Mathematics\, Physics an
d Machine Learning (IST\, Lisbon)\n\n\nAbstract\nWe compare the complexity
of training classical and quantum machine learning (ML) models for predic
ting outcomes of physical experiments. The experiments depend on an input
parameter x and involve the execution of a (possibly unknown) quantum proc
ess $E$. Our figure of merit is the number of runs of $E$ needed during tr
aining\, disregarding other measures of complexity. A classical ML perform
s a measurement and records the classical outcome after each run of $E$\,
while a quantum ML can access $E$ coherently to acquire quantum data\; the
classical or quantum data is then used to predict outcomes of future expe
riments. We prove that\, for any input distribution $D(x)$\, a classical M
L can provide accurate predictions on average by accessing $E$ a number of
times comparable to the optimal quantum ML. In contrast\, for achieving a
ccurate prediction on all inputs\, we show that exponential quantum advant
age exists in certain tasks. For example\, to predict expectation values o
f all Pauli observables in an $n-$qubit system\, we present a quantum ML u
sing only $O(n)$ data and prove that a classical ML requires $2^{\\Omega(n
)}$ data.\n
LOCATION:https://researchseminars.org/talk/MPML/36/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Mário Figueiredo (Instituto Superior Técnico and IT)
DTSTART;VALUE=DATE-TIME:20210217T180000Z
DTEND;VALUE=DATE-TIME:20210217T190000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/37
DESCRIPTION:Title: De
aling with Correlated Variables in Supervised Learning\nby Mário Figu
eiredo (Instituto Superior Técnico and IT) as part of Mathematics\, Physi
cs and Machine Learning (IST\, Lisbon)\n\n\nAbstract\nLinear (and generali
zed linear) regression (LR) is an old\, but still essential\, statistical
tool: its goal is to learn to predict a (response) variable from a linear
combination of other (explanatory) variables. A central problem in LR is t
he selection of relevant variables\, because using fewer variables tends t
o yield better generalization and because this identification may be meani
ngful (e.g.\, which genes are relevant to predict a certain disease). In t
he past quarter-century\, variable selection (VS) based on sparsity-induci
ng regularizers has been a central paradigm\, the most famous example bein
g the LASSO\, which has been intensively studied\,\nextended\, and applied
.\n\nIn many contexts\, it is natural to have highly-correlated variables
(e.g.\, several genes that are strongly co-regulated)\, thus simultaneousl
y relevant as predictors. In this case\, sparsity-based VS may fail: it ma
y select an arbitrary subset of these variables and it is unstable. Moreov
er\, it is often desirable to identify all the relevant variables\, not ju
st an arbitrary subset thereof\, a goal for which several approaches have
been proposed. This talk will be devoted to a recent class of such approac
hes\, called ordered weighted l1 (OWL). The key feature of OWL is that it
is provably able to explicitly identify (i.e. cluster) sufficiently-correl
ated features\, without having to compute these correlations. Several theo
retical results characterizing OWL will be presented\, including connectio
ns to the mathematics of economic inequality. Computational and optimizati
on aspects will also be addressed\, as well as recent applications in subs
pace clustering\, learning Gaussian graphical models\, and deep neural net
works.\n
LOCATION:https://researchseminars.org/talk/MPML/37/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Maciej Koch-J8anusz (University of Zurich)
DTSTART;VALUE=DATE-TIME:20210222T170000Z
DTEND;VALUE=DATE-TIME:20210222T180000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/38
DESCRIPTION:Title: St
atistical physics through the lens of real-space mutual information\nb
y Maciej Koch-J8anusz (University of Zurich) as part of Mathematics\, Phys
ics and Machine Learning (IST\, Lisbon)\n\n\nAbstract\nIdentifying the rel
evant coarse-grained degrees of freedom in a complex physical system is a
key stage in developing effective theories. The renormalization group (RG)
provides a framework for this task\, but its practical execution in unfam
iliar systems is fraught with ad hoc choices. Machine learning approaches\
, on the other hand\, though promising\, often lack formal interpretabilit
y: it is unclear what relation\, if any\, the architecture- and training-d
ependent learned "relevant" features bear to standard objects of physical
theory.\n\nI will present recent results addressing both issues. We develo
p a fast algorithm\, the RSMI-NE\, employing state-of-art results in machi
ne-learning-based estimation of information-theoretic quantities to constr
uct the optimal coarse-graining. We use it to develop a new approach to id
entifying the most relevant field theory operators describing a statistica
l system\, which we validate on the example of interacting dimer model. I
will also discuss formal results underlying the method: we establish equiv
alence between the information-theoretic notion of relevance defined in th
e Information Bottleneck (IB) formalism of compression theory\, and the fi
eld-theoretic relevance of the RG. We show analytically that for statistic
al physical systems the "relevant" degrees of freedom found using IB compr
ession indeed correspond to operators with the lowest scaling dimensions\,
providing a dictionary connecting two distinct theoretical toolboxes.\n
LOCATION:https://researchseminars.org/talk/MPML/38/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Markus Heyl (Max-Planck Institute for the Physics of Complex Syste
ms\, Dresden)
DTSTART;VALUE=DATE-TIME:20210322T170000Z
DTEND;VALUE=DATE-TIME:20210322T180000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/39
DESCRIPTION:Title: Qu
antum many-body dynamics in two dimensions with artificial neural networks
\nby Markus Heyl (Max-Planck Institute for the Physics of Complex Syst
ems\, Dresden) as part of Mathematics\, Physics and Machine Learning (IST\
, Lisbon)\n\n\nAbstract\nIn the last two decades the field of nonequilibri
um quantum many-body physics has seen a rapid development driven\, in part
icular\, by the remarkable progress in quantum simulators\, which today pr
ovide access to dynamics in quantum matter with an unprecedented control.
However\, the efficient numerical simulation of nonequilibrium real-time e
volution in isolated quantum matter still remains a key challenge for curr
ent computational methods especially beyond one spatial dimension. In this
talk I will present a versatile and efficient machine learning inspired a
pproach. I will first introduce the general idea of encoding quantum many-
body wave functions into artificial neural networks. I will then identify
and resolve key challenges for the simulation of real-time evolution\, whi
ch previously imposed significant limitations on the accurate description
of large systems and long-time dynamics. As a concrete example\, I will co
nsider the dynamics of the paradigmatic two-dimensional transverse field I
sing model\, where we observe collapse and revival oscillations of ferroma
gnetic order and demonstrate that the reached time scales are comparable t
o or exceed the capabilities of state-of-the-art tensor network methods.\n
LOCATION:https://researchseminars.org/talk/MPML/39/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Pedro A. Santos (Instituto Superior Técnico and INESC-ID)
DTSTART;VALUE=DATE-TIME:20210409T130000Z
DTEND;VALUE=DATE-TIME:20210409T140000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/40
DESCRIPTION:Title: Tw
o-time scale stochastic approximation for reinforcement learning with line
ar function approximation\nby Pedro A. Santos (Instituto Superior Téc
nico and INESC-ID) as part of Mathematics\, Physics and Machine Learning (
IST\, Lisbon)\n\n\nAbstract\nIn this presentation\, I will introduce some
traditional Reinforcement Learning problems and algorithms\, and analyze h
ow some problems can be avoided and convergence results obtained using a t
wo-time scale variation of the usual stochastic approximation approach.\n\
nThis variation was inspired by the practical successes of Deep Q-Learning
in attaining superhuman performance at some classical Atari games by Deep
mind's research team in 2015. Machine Learning practical successes like th
is often have no corresponding explaining theory. The work that will be pr
esented intends to contribute to that goal.\n\nJoint work with Diogo Carva
lho and Francisco Melo from INESC-ID.\n
LOCATION:https://researchseminars.org/talk/MPML/40/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Rebecca Willett (University of Chicago)
DTSTART;VALUE=DATE-TIME:20210507T130000Z
DTEND;VALUE=DATE-TIME:20210507T140000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/41
DESCRIPTION:Title: Ma
chine Learning and Inverse Problems: Deeper and More Robust\nby Rebecc
a Willett (University of Chicago) as part of Mathematics\, Physics and Mac
hine Learning (IST\, Lisbon)\n\nInteractive livestream: https://videoconf-
colibri.zoom.us/j/91599759679\nPassword hint: Register via URL in the comm
ents at least 2h before\n\nAbstract\nMany challenging image processing tas
ks can be described by an ill-posed linear inverse problem: deblurring\, d
econvolution\, inpainting\, compressed sensing\, and superresolution all l
ie in this framework. Recent advances in machine learning and image proces
sing have illustrated that it is often possible to learn a regularizer fro
m training data that can outperform more traditional approaches by large m
argins. In this talk\, I will describe the central prevailing themes of th
is emerging area and a taxonomy that can be used to categorize different p
roblems and reconstruction methods. We will also explore mechanisms for mo
del adaptation\; that is\, given a network trained to solve an initial inv
erse problem with a known forward model\, we propose novel procedures that
adapt the network to a perturbed forward model\, even without full knowle
dge of the perturbation. Finally\, I will describe a new class of approach
es based on "infinite-depth networks" that can yield up to a 4dB PSNR impr
ovement in reconstruction accuracy above state-of-the-art alternatives and
where the computational budget can be selected at test time to optimize c
ontext-dependent trade-offs between accuracy and computation.\n
LOCATION:https://researchseminars.org/talk/MPML/41/
URL:https://videoconf-colibri.zoom.us/j/91599759679
END:VEVENT
BEGIN:VEVENT
SUMMARY:Mikhail Belkin (Halicioğlu Data Science Institute\, University of
California San Diego)
DTSTART;VALUE=DATE-TIME:20210428T170000Z
DTEND;VALUE=DATE-TIME:20210428T180000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/42
DESCRIPTION:by Mikhail Belkin (Halicioğlu Data Science Institute\, Univer
sity of California San Diego) as part of Mathematics\, Physics and Machine
Learning (IST\, Lisbon)\n\nInteractive livestream: https://videoconf-coli
bri.zoom.us/j/91599759679\nPassword hint: Register via URL in the comments
at least 2h before\nAbstract: TBA\n
LOCATION:https://researchseminars.org/talk/MPML/42/
URL:https://videoconf-colibri.zoom.us/j/91599759679
END:VEVENT
BEGIN:VEVENT
SUMMARY:Gabriel Peyré (École Normale Supérieure)
DTSTART;VALUE=DATE-TIME:20210414T170000Z
DTEND;VALUE=DATE-TIME:20210414T180000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/43
DESCRIPTION:Title: Sc
aling Optimal Transport for High dimensional Learning\nby Gabriel Peyr
é (École Normale Supérieure) as part of Mathematics\, Physics and Machi
ne Learning (IST\, Lisbon)\n\n\nAbstract\nOptimal transport (OT) has recen
tly gained lot of interest in machine learning. It is a natural tool to co
mpare in a geometrically faithful way probability distributions. It finds
applications in both supervised learning (using geometric loss functions)
and unsupervised learning (to perform generative model fitting). OT is how
ever plagued by the curse of dimensionality\, since it might require a num
ber of samples which grows exponentially with the dimension. In this talk\
, I will explain how to leverage entropic regularization methods to define
computationally efficient loss functions\, approximating OT with a better
sample complexity.\n\nMore information and references can be found on the
website of our book\n"Computational Optimal Transport"\, https://optimalt
ransport.github.io/\n
LOCATION:https://researchseminars.org/talk/MPML/43/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Kyriakos Vamvoudakis (Georgia Institute of Technology)
DTSTART;VALUE=DATE-TIME:20210521T130000Z
DTEND;VALUE=DATE-TIME:20210521T140000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/44
DESCRIPTION:by Kyriakos Vamvoudakis (Georgia Institute of Technology) as p
art of Mathematics\, Physics and Machine Learning (IST\, Lisbon)\n\nIntera
ctive livestream: https://videoconf-colibri.zoom.us/j/91599759679\nPasswor
d hint: Register via URL in the comments at least 2h before\nAbstract: TBA
\n
LOCATION:https://researchseminars.org/talk/MPML/44/
URL:https://videoconf-colibri.zoom.us/j/91599759679
END:VEVENT
BEGIN:VEVENT
SUMMARY:Gustau Camps-Valls (Universitat de València)
DTSTART;VALUE=DATE-TIME:20210528T130000Z
DTEND;VALUE=DATE-TIME:20210528T140000Z
DTSTAMP;VALUE=DATE-TIME:20210419T095018Z
UID:MPML/45
DESCRIPTION:Title: Ph
ysics Aware Machine Learning for the Earth Sciences\nby Gustau Camps-V
alls (Universitat de València) as part of Mathematics\, Physics and Machi
ne Learning (IST\, Lisbon)\n\nInteractive livestream: https://videoconf-co
libri.zoom.us/j/91599759679\nPassword hint: Register via URL in the commen
ts at least 2h before\n\nAbstract\nMost problems in Earth sciences aim to
do inferences about the system\, where accurate predictions are just a tin
y part of the whole problem. Inferences mean understanding variables relat
ions\, deriving models that are physically interpretable\, that are simple
parsimonious\, and mathematically tractable. Machine learning models alon
e are excellent approximators\, but very often do not respect the most ele
mentary laws of physics\, like mass or energy conservation\, so consistenc
y and confidence are compromised. I will review the main challenges ahead
in the field\, and introduce several ways to live in the Physics and machi
ne learning interplay that allows us (1) to encode differential equations
from data\, (2) constrain data-driven models with physics-priors and depen
dence constraints\, (3) improve parameterizations\, (4) emulate physical m
odels\, and (5) blend data-driven and process-based models. This is a coll
ective long-term AI agenda towards developing and applying algorithms capa
ble of discovering knowledge in the Earth system.\n
LOCATION:https://researchseminars.org/talk/MPML/45/
URL:https://videoconf-colibri.zoom.us/j/91599759679
END:VEVENT
END:VCALENDAR