BEGIN:VCALENDAR
VERSION:2.0
PRODID:researchseminars.org
CALSCALE:GREGORIAN
X-WR-CALNAME:researchseminars.org
BEGIN:VEVENT
SUMMARY:Jason Weston (Facebook)
DTSTART:20201005T190000Z
DTEND:20201005T200000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/1
DESCRIPTION:Title: LIGHT: Training Agents that can Act and Speak with Other Models a
nd Humans in a Rich Text Adventure Game World\nby Jason Weston (Facebo
ok) as part of Machine Learning Advances and Applications Seminar\n\n\nAbs
tract\nThe title just about covered it\, but.. LIGHT is a rich fantasy tex
t adventure game environment featuring dialogue and actions between agents
in the world\, which consist of both models and humans.\n\nI will summari
ze work on this platform\, including crowdsourcing and machine learning to
build the rich world environment\, training agents to speak and act withi
n it\, and deploying the game for lifelong learning of agents by interacti
ng with humans.\n\nThis is joint work with a number of authors: Emily Dina
n\, Angela Fan\, Samuel Humeau\, Saachi Jain\, Siddharth Karamcheti\, Douw
e Kiela\, Margaret Li\, Shrimai Prabhumoye\, Emma Qian\, Tim Rocktäschel\
, Pratik Ringshia\, Kurt Shuster\, Arthur Szlam\, Adina Williams. See http
s://parl.ai/projects/light/ for papers & more info!\n\nBio: Jason Weston i
s a Research Scientist at Facebook\, NY and a Visiting Research Professor
at NYU. He earned his PhD in Machine Learning at Royal Holloway\, Universi
ty of London and at AT&T Research in Red Bank\, NJ (advisors: Alex Gammerm
an\, Volodya Vovk and Vladimir Vapnik) in 2000. From 2000 to 2001\, he was
a Researcher at Biowulf Technologies. From 2002 to 2003 he was a Research
Scientist at the Max Planck Institute for Biological Cybernetics in Tuebi
ngen\, Germany. From 2003 to 2009 he was a research staff member at NEC La
bs America\, Princeton. From 2009 to 2014 he was a Research Scientist at G
oogle\, NY. His interests lie in statistical machine learning\, with a foc
us on reasoning\, memory\, perception\, interaction and communication. Jas
on has published over 100 papers\, including best paper awards at ICML and
ECML\, and a Test of Time Award for his work "A Unified Architecture for
Natural Language Processing: Deep Neural Networks with Multitask Learning\
," ICML 2008 (with Ronan Collobert). He was part of the YouTube team that
won a National Academy of Television Arts & Sciences Emmy Award for Techno
logy and Engineering for Personalized Recommendation Engines for Video Dis
covery. He was listed as the 16th most influential machine learning schola
r at AMiner and one of the top 50 authors in Computer Science in Science.\
n
LOCATION:https://researchseminars.org/talk/Machine_Learning/1/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Sham Kakade (University of Washington)
DTSTART:20201102T200000Z
DTEND:20201102T210000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/2
DESCRIPTION:Title: Policy Gradient Methods\, Curvature\, and Distribution Shift\
nby Sham Kakade (University of Washington) as part of Machine Learning Adv
ances and Applications Seminar\n\n\nAbstract\nReinforcement learning is no
w the dominant paradigm for how an agent learns to interact with the world
in order to achieve some long term objectives. Here\, policy gradient met
hods are among the most effective methods in challenging reinforcement lea
rning problems\, due to that they: are applicable to any differentiable po
licy parameterization\; admit easy extensions to function approximation\;
easily incorporate structured state and action spaces\; are easy to implem
ent in a simulation based\, model-free manner.\n\nHowever\, little is know
n about even their most basic theoretical convergence properties\, includi
ng:\n\n- do they converge to a globally optimal solution\, say with a suff
iciently rich policy class?\n\n- how well do they cope with approximation
error\, say due to using a class of neural policies?\n\n- what is their fi
nite sample complexity?\n\nThis talk will survey a number of results on th
ese basic questions. We will highlight the interplay of theory\, algorithm
design\, and practice.\n\nJoint work with: Alekh Agarwal\, Jason Lee\, Ga
urav Mahajan\n\nBio: Sham Kakade is a professor in the Department of Compu
ter Science and the Department of Statistics at the University of Washingt
on and is also a senior principal researcher at Microsoft Research. His wo
rk is on the mathematical foundations of machine learning and AI. Sham's t
hesis helped lay the statistical foundations of reinforcement learning. Wi
th his collaborators\, his additional contributions include: one of the fi
rst provably efficient policy search methods in reinforcement learning\; d
eveloping the mathematical foundations for the widely used linear bandit m
odels and the Gaussian process bandit models\; the tensor and spectral met
hodologies for provable estimation of latent variable models\; the first s
harp analysis of the perturbed gradient descent algorithm\, along with the
design and analysis of numerous other convex and non-convex algorithms. H
e is the recipient of the ICML Test of Time Award\, the IBM Pat Goldberg b
est paper award\, and INFORMS Revenue Management and Pricing Prize. He has
been program chair for COLT 2011.\n\nSham was an undergraduate at Caltech
\, where he studied physics and worked under the guidance of John Preskill
in quantum computing. He completed his Ph.D. with Peter Dayan in computat
ional neuroscience at the Gatsby Unit at University College London. He was
a postdoc with Michael Kearns at the University of Pennsylvania.\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/2/
END:VEVENT
BEGIN:VEVENT
SUMMARY:David Lopez-Paz (Facebook)
DTSTART:20201130T200000Z
DTEND:20201130T210000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/3
DESCRIPTION:by David Lopez-Paz (Facebook) as part of Machine Learning Adva
nces and Applications Seminar\n\nAbstract: TBA\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/3/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Douglas Eck (Google)
DTSTART:20201214T200000Z
DTEND:20201214T210000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/4
DESCRIPTION:Title: Challenges in Building ML Algorithms for the Creative Community\nby Douglas Eck (Google) as part of Machine Learning Advances and Appli
cations Seminar\n\n\nAbstract\nMagenta is an open-source project exploring
the role of machine learning as a tool in the creative process. We've bee
n running in public (g.co/magenta) for over four years. This talk will loo
k back at successes and frustrations in bringing our work to creators\, mo
stly musicians. I'll also talk about some current and future work. Magenta
is made up of several ML researchers and engineers on the Google Brain te
am\, which focuses on deep learning. Our successes have mostly been in the
area of new algorithm development (NSynth\, MusicVAE\, Music Transformer\
, DDSP and others). Our frustrations have been in finding ways to make the
se models useful for music creators. The talk will be a casual example-dri
ven discussion about what worked and what didn't\, and where we're going n
ext. Spoiler: we have been humbled by the user interface challenges encoun
tered when building tools for creative work. My main message for Vector In
stitute is that machine learning alone is not enough to address a challeng
e like enabling new forms of creativity -- you need to think about what ar
tists really want and how to communicate with them.\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/4/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Moritz Hardt (UC Berkeley)
DTSTART:20210125T200000Z
DTEND:20210125T210000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/6
DESCRIPTION:Title: Performative Prediction\nby Moritz Hardt (UC Berkeley) as par
t of Machine Learning Advances and Applications Seminar\n\n\nAbstract\nWhe
n predictive models support decisions they can influence the outcome they
aim to predict. We call such predictions performative\; the prediction inf
luences the target. Performativity is a well-studied phenomenon in policy-
making that has so far been neglected in supervised learning. When ignored
\, performativity surfaces as undesirable distribution shift\, routinely a
ddressed with retraining. \n\nIn this talk\, I will describe a risk minimi
zation framework for performative prediction bringing together concepts fr
om statistics\, game theory\, and causality. A new element is an equilibri
um notion called performative stability. Performative stability implies th
at the predictions are calibrated not against past outcomes\, but against
the future outcomes that manifest from acting on the prediction. \n\nI wil
l then discuss recent results on performative prediction including necessa
ry and sufficient conditions for the convergence of retraining to a perfor
matively stable point of nearly minimal loss. \n\nJoint work with Juan C.
Perdomo\, Tijana Zrnic\, and Celestine Mendler-Dünner.\n\nBio: Moritz Har
dt is an Assistant Professor in the Department of Electrical Engineering a
nd Computer Sciences at the University of California\, Berkeley. Hardt inv
estigates algorithms and machine learning with a focus on reliability\, va
lidity\, and societal impact. After obtaining a PhD in Computer Science fr
om Princeton University\, he held positions at IBM Research Almaden\, Goog
le Research and Google Brain. Hardt is a co-founder of the Workshop on Fai
rness\, Accountability\, and Transparency in Machine Learning (FAT/ML) and
a co-author of the forthcoming textbook "Fairness and Machine Learning".
He has received an NSF CAREER award\, a Sloan fellowship\, and best paper
awards at ICML 2018 and ICLR 2017.\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/6/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Peter Dayan (Max Planck Institute)
DTSTART:20210208T200000Z
DTEND:20210208T210000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/7
DESCRIPTION:Title: Peril\, Prudence and Planning as Risk\, Avoidance and Worry\n
by Peter Dayan (Max Planck Institute) as part of Machine Learning Advances
and Applications Seminar\n\n\nAbstract\nRisk occupies a central role in b
oth the theory and practice of decision-making. Although it is deeply impl
icated in many conditions involving dysfunctional behavior and thought\, m
odern theoretical approaches to understanding and mitigating risk in eithe
r one-shot or sequential settings have yet to permeate fully the fields of
neural reinforcement learning and computational psychiatry. I will discus
s the use of one prominent approach\, called conditional value-at-risk to
examine both the nature of risk avoidant choices\, encompassing such thing
s as justified gambler's fallacies\, and the optimal planning that can lea
d to consideration of such choices\, with implications for offline\, rumin
ative\, thinking in the context of anxiety.\n\nThis is joint work with Chr
is Gagne.\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/7/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Chelsea Finn (Stanford/Google)
DTSTART:20210222T200000Z
DTEND:20210222T210000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/8
DESCRIPTION:Title: Principles for Tackling Distribution Shift: Pessimism\, Adaptatio
n\, and Anticipation\nby Chelsea Finn (Stanford/Google) as part of Mac
hine Learning Advances and Applications Seminar\n\n\nAbstract\nWhile we ha
ve seen substantial progress in machine learning\, a critical shortcoming
of current methods lies in handling distribution shift between training an
d deployment. Distribution shift is pervasive in real-world problems rangi
ng from natural variation in the distribution over locations or domains\,
to shifts in the distribution arising from different decision making polic
ies\, to shifts over time as the world changes. In this talk\, I'll discus
s three general principles for tackling these forms of distribution shift:
pessimism\, adaptation\, and anticipation. I'll present the most general
form of each principle before providing concrete instantiations of using e
ach in practice. This will include a simple method for substantially impro
ving robustness to spurious correlations\, a framework for quickly adaptin
g a model to a new user or domain with only unlabeled data\, and an algori
thm that enables robots to anticipate and adapt to shifts caused by other
agents.\n\nBio: Chelsea Finn is an Assistant Professor in Computer Science
and Electrical Engineering at Stanford University. Finn's research intere
sts lie in the capability of robots and other agents to develop broadly in
telligent behavior through learning and interaction. To this end\, her wor
k has included deep learning algorithms for concurrently learning visual p
erception and control in robotic manipulation skills\, self-supervised met
hods for learning a breadth of vision-based control tasks\, and meta-learn
ing algorithms that can enable fast\, few-shot adaptation in both visual p
erception and deep reinforcement learning. Finn received her Bachelor's de
gree in Electrical Engineering and Computer Science at MIT and her PhD in
Computer Science at UC Berkeley. Her research has been recognized through
the Microsoft Research Faculty Fellowship\, the ACM doctoral dissertation
award\, the C.V. Ramamoorthy Distinguished Research Award\, and the MIT Te
chnology Review 35 under 35 Award\, and her work has been covered by vario
us media outlets\, including the New York Times\, Wired\, and Bloomberg.\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/8/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Percy Liang (Stanford Universi)
DTSTART:20210308T200000Z
DTEND:20210308T210000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/9
DESCRIPTION:Title: Deep Models and on Shaping their Development\nby Percy Liang
(Stanford Universi) as part of Machine Learning Advances and Applications
Seminar\n\n\nAbstract\nModels in deep learning are wild beasts: they devou
r raw data\, are powerful but hard to control. This talk explores two appr
oaches to taming them. First\, I will introduce concept bottleneck network
s\, in which a deep neural network makes a prediction via interpretable\,
high-level concepts. We show that such models can obtain comparable accura
cy with standard models\, while offering the unique ability for a human to
perform test-time interventions on the concepts. Second\, I will introduc
e prefix-tuning\, which allows one to harness the power of pre-trained lan
guage models (e.g.\, GPT-2) for text generation tasks. The key idea is to
learn a continuous task-specific prefix that primes the language model for
the task at hand. Prefix-tuning obtains comparable accuracy to fine-tunin
g\, while only updating 0.1% of the parameters. Finally\, I will end with
a broad question: what kind of datasets should the community develop to dr
ive innovation in modeling approaches? Are size and realism necessary attr
ibutes of a dataset? Could we have made all the modeling progress in NLP w
ithout SQuAD? As this counterfactual question is impossible to answer\, we
perform a retrospective study on 20 modeling approaches and show that eve
n a small\, synthetic dataset can track the progress that was made on SQuA
D. While inconclusive\, this result encourages us to think more critically
about the value of datasets during their construction.\n\nBio: Percy Lian
g is an Associate Professor of Computer Science at Stanford University (B.
S. from MIT\, 2004\; Ph.D. from UC Berkeley\, 2011). His research spans ma
ny topics in machine learning and natural language processing\, including
robustness\, interpretability\, semantics\, and reasoning. He is also a st
rong proponent of reproducibility through the creation of CodaLab Workshee
ts. His awards include the Presidential Early Career Award for Scientists
and Engineers (2019)\, IJCAI Computers and Thought Award (2016)\, an NSF C
AREER Award (2016)\, a Sloan Research Fellowship (2015)\, a Microsoft Rese
arch Faculty Fellowship (2014)\, and multiple paper awards at ACL\, EMNLP\
, ICML\, and COLT.\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/9/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Daphne Koller (Stanford University)
DTSTART:20210322T190000Z
DTEND:20210322T200000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/10
DESCRIPTION:Title: Machine Learning: A new approach to drug discovery\nby Daphn
e Koller (Stanford University) as part of Machine Learning Advances and Ap
plications Seminar\n\n\nAbstract\nModern medicine has given us effective t
ools to treat some of the most significant and burdensome diseases. At the
same time\, it is becoming consistently more challenging and more expensi
ve to develop new therapeutics. A key factor in this trend is that the dru
g development process involves multiple steps\, each of which involves a c
omplex and protracted experiment that often fails. We believe that\, for m
any of these phases\, it is possible to develop machine learning models to
help predict the outcome of these experiments\, and that those models\, w
hile inevitably imperfect\, can outperform predictions based on traditiona
l heuristics. To achieve this goal\, we are bringing together high-quality
data from human cohorts\, while also developing cutting edge methods in h
igh throughput biology and chemistry that can produce massive amounts of i
n vitro data relevant to human disease and therapeutic interventions. Thos
e are then used to train machine learning models that make predictions abo
ut novel targets\, coherent patient segments\, and the clinical effect of
molecules. Our ultimate goal is to develop a new approach to drug developm
ent that uses high-quality data and ML models to design novel\, safe\, and
effective therapies that help more people\, faster\, and at a lower cost.
\n\nBio: Daphne Koller is CEO and Founder of insitro\, a machine-learning
enabled drug discovery company. Daphne is also co-founder of Engageli\, wa
s the Rajeev Motwani Professor of Computer Science at Stanford University\
, where she served on the faculty for 18 years\, the co-CEO and President
of Coursera\, and the Chief Computing Officer of Calico\, an Alphabet comp
any in the healthcare space. She is the author of over 200 refereed public
ations appearing in venues such as Science\, Cell\, and Nature Genetics. D
aphne was recognized as one of TIME Magazine's 100 most influential people
in 2012. She received the MacArthur Foundation Fellowship in 2004 and the
ACM Prize in Computing in 2008. She was inducted into the National Academ
y of Engineering in 2011 and elected a fellow of the American Association
for Artificial Intelligence in 2004\, the American Academy of Arts and Sci
ences in 2014\, and the International Society of Computational Biology in
2017.\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/10/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Surya Ganguli (Stanford University)
DTSTART:20210405T190000Z
DTEND:20210405T200000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/11
DESCRIPTION:Title: Weaving together machine learning\, theoretical physics\, and ne
uroscience\nby Surya Ganguli (Stanford University) as part of Machine
Learning Advances and Applications Seminar\n\n\nAbstract\nAn exciting area
of intellectual activity in this century may well revolve around a synthe
sis of machine learning\, theoretical physics\, and neuroscience. The unif
ication of these fields will likely enable us to exploit the power of comp
lex systems analysis\, developed in theoretical physics and applied mathem
atics\, to elucidate the design principles governing neural systems\, both
biological and artificial\, and deploy these principles to develop better
algorithms in machine learning. We will give several vignettes in this di
rection\, including: (1) determining the best optimization problem to solv
e in order to perform regression in high dimensions\; (2) finding exact so
lutions to the dynamics of generalization error in deep linear networks\;
(3) derving the detailed structure of the primate retina by analyzing opti
mal convolutional auto-encoders of natural movies\; (4) analyzing and expl
aining the origins of hexagonal firing patterns in recurrent neural networ
ks trained to path-integrate\; (5) understanding the geometry and dynamics
of high dimensional optimization in the classical limit of dissipative ma
ny-body quantum optimizers.\n\nReferences:\n\nM. Advani and S. Ganguli\, S
tatistical mechanics of optimal convex inference in high dimensions\, Phys
ical Review X\, 6\, 031034\, 2016.\n\nM. Advani and S. Ganguli\, An equiva
lence between high dimensional Bayes optimal inference and M-estimation\,
NeurIPS\, 2016.\n\nA.K. Lampinen and S. Ganguli\, An analytic theory of ge
neralization dynamics and transfer learning in deep linear networks\, Inte
rnational Conference on Learning Representations (ICLR)\, 2019.\n\nH. Tana
ka\, A. Nayebi\, N. Maheswaranathan\, L.M. McIntosh\, S. Baccus\, S. Gangu
li\, From deep learning to mechanistic understanding in neuroscience: the
structure of retinal prediction\, NeurIPS 2019.\n\nS. Deny\, J. Lindsey\,
S. Ganguli\, S. Ocko\, The emergence of multiple retinal cell types throug
h efficient coding of natural movies\, Neural Information Processing Syste
ms (NeurIPS) 2018.\n\nB. Sorscher\, G. Mel\, S. Ganguli\, S. Ocko\, A unif
ied theory for the origin of grid cells through the lens of pattern format
ion\, NeurIPS 2019.\n\nY. Bahri\, J. Kadmon\, J. Pennington\, S. Schoenhol
z\, J. Sohl-Dickstein\, and S. Ganguli\, Statistical mechanics of deep lea
rning\, Annual Reviews of Condensed Matter Physics\, 2020.\n\nY. Yamamoto\
, T. Leleu\, S. Ganguli and H. Mabuchi\, Coherent Ising Machines: quantum
optics and neural network perspectives\, Applied Physics Letters 2020.\n\n
B.P. Marsh\, Y\, Guo\, R.M. Kroeze\, S. Gopalakrishnan\, S. Ganguli\, J. K
eeling\, B.L. Lev\n\nEnhancing associative memory recall and storage capac
ity using confocal cavity QED\, https://arxiv.org/abs/2009.01227.\n\nBio:
Surya Ganguli triple majored in physics\, mathematics\, and EECS at MIT\,
completed a PhD in string theory at Berkeley\, and a postdoc in theoretica
l neuroscience at UCSF. He is now an associate professor of Applied physic
s at Stanford where he leads the Neural Dynamics and Computation Lab. His
research spans the fields of neuroscience\, machine learning and physics\,
focusing on understanding and improving how both biological and artificia
l neural networks learn striking emergent computations. He has been awarde
d a Swartz-Fellowship in computational neuroscience\, a Burroughs-Wellcome
Career Award\, a Terman Award\, a NeurIPS Outstanding Paper Award\, a Slo
an fellowship\, a James S. McDonnell Foundation scholar award in human cog
nition\, a McKnight Scholar award in Neuroscience\, a Simons Investigator
Award in the mathematical modeling of living systems\, and an NSF career a
ward.\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/11/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Jakob Foerster (University of Toronto and Vector Institute)
DTSTART:20210111T200000Z
DTEND:20210111T203000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/12
DESCRIPTION:Title: Zero-Shot (Human-AI) Coordination (in Hanabi) and Ridge Rider\nby Jakob Foerster (University of Toronto and Vector Institute) as part
of Machine Learning Advances and Applications Seminar\n\n\nAbstract\nIn re
cent years we have seen fast progress on a number of zero-sum benchmark pr
oblems in AI\, e.g. Go\, Poker and Dota. In contrast\, success in the real
world requires humans to collaborate and communicate with others\, in set
tings that are\, at least partially\, cooperative. Recently\, the card gam
e Hanabi has been established as a new benchmark environment to fill this
gap. In particular\, Hanabi is interesting to humans since it is entirely
focused on theory of mind\, i.e.\, the ability to reason over the intentio
ns\, beliefs and point of view of other agents when observing their action
s. This is particularly important in applications such as communication\,
assistive technologies and autonomous driving.\n\nWe start out by introduc
ing the zero-shot coordination setting as a new frontier for multi-agent r
esearch\, which is partially addressed by Other-Play\, a novel learning al
gorithm which biases learning towards more human compatible policies.\n\nL
astly we introduce Ridge Rider\, our brand new algorithm which addresses b
oth zero-shot coordination and other optimization problems where the objec
tive we care about can by definition not be evaluated during training time
.\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/12/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Shakir Mohamed (DeepMind)
DTSTART:20210419T190000Z
DTEND:20210419T200000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/13
DESCRIPTION:Title: AI and Shared Prosperity\nby Shakir Mohamed (DeepMind) as pa
rt of Machine Learning Advances and Applications Seminar\n\n\nAbstract\nTh
e question we'll explore in this talk is how we can redirect the path we a
re taking as AI designers towards the promotion of shared prosperity\, and
in designing and deploying AI in ways where both the benefits\, and the r
isks\, of new technology are shared equally across society. I'll explore t
hese ideas by combining different elements of my recent work. I'll start w
ith a concrete problem of developing AI for detecting organ damage in hosp
itals. This use-case will highlight the interconnected sociotechnical syst
em that machine learning operates within. This system has sets of values a
nd politics that must contend with a colonial legacy and coloniality\, and
I'll explore thinking on decolonial AI\, and also delve into a further us
e case on queer fairness. Along the way\, I'll try to connect to the work
of other organistions\, like the AI and Shared Prosperity initiative by th
e Partnership on AI\, and Royal Society programme on digital technologies
for the planet.\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/13/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Chris Maddison (University of Toronto and Vector Institute)
DTSTART:20210111T203000Z
DTEND:20210111T210000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/14
DESCRIPTION:Title: Gradient Estimation with Stochastic Softmax Tricks\nby Chris
Maddison (University of Toronto and Vector Institute) as part of Machine
Learning Advances and Applications Seminar\n\n\nAbstract\nGradient estimat
ion is an important problem in modern machine learning frameworks that rel
y heavily on gradient-based optimization. For gradient estimation in the p
resence of discrete random variables\, the Gumbel-based relaxed gradient e
stimators are easy to implement and low variance\, but the goal of scaling
them comprehensively to large combinatorial distributions is still outsta
nding. Working within the perturbation model framework\, we introduce stoc
hastic softmax tricks\, which generalize the Gumbel-Softmax trick to combi
natorial spaces. Our framework is a unified perspective on existing relaxe
d estimators for perturbation models\, and it contains many novel relaxati
ons. We design structured relaxations for subset selection\, spanning tree
s\, arborescences\, and others. We consider an application to helping make
machine learning models more explainable.\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/14/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Courtney Paquette (McGill University)
DTSTART:20211108T200000Z
DTEND:20211108T210000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/15
DESCRIPTION:Title: SGD in the Large: Average-case Analysis\, Asymptotics\, and Step
size Criticality\nby Courtney Paquette (McGill University) as part of
Machine Learning Advances and Applications Seminar\n\n\nAbstract\nIn this
talk\, I will present a framework\, inspired by random matrix theory\, for
analyzing the dynamics of stochastic gradient descent (SGD) when both the
number of samples and dimensions are large. Using this new framework\, we
show that the dynamics of SGD on a least squares problem with random data
become deterministic in the large sample and dimensional limit. Furthermo
re\, the limiting dynamics are governed by a Volterra integral equation. T
his model predicts that SGD undergoes a phase transition at an explicitly
given critical stepsize that ultimately affects its convergence rate\, whi
ch we also verify experimentally. Finally\, when input data is isotropic\,
we provide explicit expressions for the dynamics and average-case converg
ence rates. These rates show significant improvement over the worst-case c
omplexities.\n\nBio: Courtney Paquette is an assistant professor at McGill
University and a CIFAR Canada AI chair\, MILA. Paquette's research broadl
y focuses on designing and analyzing algorithms for large-scale optimizati
on problems\, motivated by applications in data science. She received her
PhD from the mathematics department at the University of Washington (2017)
\, held postdoctoral positions at Lehigh University (2017-2018) and Univer
sity of Waterloo (NSF postdoctoral fellowship\, 2018-2019)\, and was a res
earch scientist at Google Research\, Brain Montreal (2019-2020).\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/15/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Lester Mackey (Microsoft/Stanford University)
DTSTART:20211122T200000Z
DTEND:20211122T210000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/16
DESCRIPTION:Title: Kernel Thinning and Stein Thinning\nby Lester Mackey (Micros
oft/Stanford University) as part of Machine Learning Advances and Applicat
ions Seminar\n\n\nAbstract\nThis talk will introduce two new tools for sum
marizing a probability distribution more effectively than independent samp
ling or standard Markov chain Monte Carlo thinning:\n\n1. Given an initial
n point summary (for example\, from independent sampling or a Markov chai
n)\, kernel thinning finds a subset of only square-root n points with comp
arable worst-case integration error across a reproducing kernel Hilbert sp
ace.\n\n2. If the initial summary suffers from biases due to off-target sa
mpling\, tempering\, or burn-in\, Stein thinning simultaneously compresses
the summary and improves the accuracy by correcting for these biases.\n\n
These tools are especially well-suited for tasks that incur substantial do
wnstream computation costs per summary point like organ and tissue modelin
g in which each simulation consumes 1000s of CPU hours.\n\nBio: Lester Mac
key is a Principal Researcher at Microsoft Research\, where he develops ma
chine learning methods\, models\, and theory for large-scale learning task
s driven by applications from climate forecasting\, healthcare\, and the s
ocial good. Lester moved to Microsoft from Stanford University\, where he
was an assistant professor of Statistics and (by courtesy) of Computer Sci
ence. He earned his PhD in Computer Science and MA in Statistics from UC B
erkeley and his BSE in Computer Science from Princeton University. He co-o
rganized the second place team in the Netflix Prize competition for collab
orative filtering\, won the Prize4Life ALS disease progression prediction
challenge\, won prizes for temperature and precipitation forecasting in th
e yearlong real-time Subseasonal Climate Forecast Rodeo\, and received bes
t paper and best student paper awards from the ACM Conference on Programmi
ng Language Design and Implementation and the International Conference on
Machine Learning.\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/16/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Andreas Krause (ETH Zürich)
DTSTART:20220124T200000Z
DTEND:20220124T210000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/17
DESCRIPTION:by Andreas Krause (ETH Zürich) as part of Machine Learning Ad
vances and Applications Seminar\n\nAbstract: TBA\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/17/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Joan Bruna Estrach (New York University)
DTSTART:20220207T200000Z
DTEND:20220207T210000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/18
DESCRIPTION:by Joan Bruna Estrach (New York University) as part of Machine
Learning Advances and Applications Seminar\n\nAbstract: TBA\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/18/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Matthew Johnson (Google)
DTSTART:20220214T200000Z
DTEND:20220214T210000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/19
DESCRIPTION:by Matthew Johnson (Google) as part of Machine Learning Advanc
es and Applications Seminar\n\nAbstract: TBA\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/19/
END:VEVENT
BEGIN:VEVENT
SUMMARY:David Pfau (DeepMind)
DTSTART:20220307T200000Z
DTEND:20220307T210000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/20
DESCRIPTION:by David Pfau (DeepMind) as part of Machine Learning Advances
and Applications Seminar\n\nAbstract: TBA\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/20/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Jan Peters (Technische Universitaet Darmstadt)
DTSTART:20220321T190000Z
DTEND:20220321T200000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/21
DESCRIPTION:by Jan Peters (Technische Universitaet Darmstadt) as part of M
achine Learning Advances and Applications Seminar\n\nAbstract: TBA\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/21/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Philipp Hennig (Max Planck Institute for Intelligent Systems)
DTSTART:20220404T190000Z
DTEND:20220404T200000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/22
DESCRIPTION:by Philipp Hennig (Max Planck Institute for Intelligent System
s) as part of Machine Learning Advances and Applications Seminar\n\nAbstra
ct: TBA\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/22/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Cynthia Rudin (Duke University)
DTSTART:20220425T190000Z
DTEND:20220425T200000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/23
DESCRIPTION:by Cynthia Rudin (Duke University) as part of Machine Learning
Advances and Applications Seminar\n\nAbstract: TBA\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/23/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Anima Anandkumar (California Institute of Technology)
DTSTART:20220502T190000Z
DTEND:20220502T200000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/24
DESCRIPTION:by Anima Anandkumar (California Institute of Technology) as pa
rt of Machine Learning Advances and Applications Seminar\n\nAbstract: TBA\
n
LOCATION:https://researchseminars.org/talk/Machine_Learning/24/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Radford Neal (University of Toronto)
DTSTART:20220516T190000Z
DTEND:20220516T200000Z
DTSTAMP:20260422T212512Z
UID:Machine_Learning/25
DESCRIPTION:by Radford Neal (University of Toronto) as part of Machine Lea
rning Advances and Applications Seminar\n\nAbstract: TBA\n
LOCATION:https://researchseminars.org/talk/Machine_Learning/25/
END:VEVENT
END:VCALENDAR