BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CERN//INDICO//EN
BEGIN:VEVENT
SUMMARY:Alan Turing Lectures
DTSTART;VALUE=DATE-TIME:20150107T082500Z
DTEND;VALUE=DATE-TIME:20150107T113000Z
DTSTAMP;VALUE=DATE-TIME:20240416T140503Z
UID:indico-event-4130@cern.ch
DESCRIPTION:Inauguration: 1:55 PM\n\nThe Alan Turing lecture series is a n
ew initiative of ICTS. In this series\, eminent Biologists\, Computer Scie
ntists\, and Engineers would be invited to deliver lectures on significant
developments in their areas. Details about this inaugural Turing lecture
series and the lecturers are given below\,\n\nProf. Sanjeev Arora\nPrincet
on University\n\nTime: 2:00 - 2:45 PM\n\nOvercoming Computational Intracta
bility in Unsupervised Learning\nGiven today’s data deluge\, unsupervise
d learning i.e. learning with unlabeled data is becoming increasingly impo
rtant. Most natural problems in this domain e.g. for models such as mixtur
e models\, HMMs\, graphical models\, topic models and sparse coding/dictio
nary learning\, deep learning are NP-hard. Therefore researchers in practi
ce use either heuristics or convex relaxations with no concrete approximat
ion bounds. Several nonconvex heuristics work well in practice\, which is
also a mystery. The talk will describe a sequence of recent results whereb
y rigorous approaches leading to polynomial running time are possible for
several problems in unsupervised learning. The proof of polynomial running
time usually relies upon nondegeneracy assumptions on the data and the mo
del parameters\, and often also on stochastic properties of the data (aver
age-case analysis). Some of these new algorithms are very efficient and pr
actical. e.g. for topic modeling.\n\n\nProf. Robert Schapire\nMicrosoft Re
search and Princeton University\n\nTime: 2:45 - 3:30 PM\n\nThe Contextual
Bandits Problem: A New\, Fast\, and Simple Algorithm\nIn the contextual ba
ndits learning problem\, the learner must repeatedly decide which action t
o take in response to an observed context\, and is then permitted to obser
ve the received reward\, but only for the chosen action. The goal is to
learn through experience to behave nearly as well as the best policy (or d
ecision rule) in some possibly very large space of candidate policies. W
e assume that the learner can only access this policy space using an oracl
e for solving empirical cost-sensitive classification problems\; in practi
ce\, most off-the-shelf classification algorithms could be used for this p
urpose. In this very general setting\, we present a new\, fast\, and sim
ple algorithm that achieves a regret guarantee that is statistically optim
al. Moreover\, this algorithm makes very modest use of the oracle\, whic
h it calls far less than once per round\, on average. These properties s
uggest this may be the most practical contextual bandits algorithm among a
ll existing approaches that are provably effective for general policy clas
ses.\nThis is joint work with Alekh Agarwal\, Daniel Hsu\, Satyen Kale\, J
ohn Langford and Lihong Li.\n\n \n\nProf. Ravi Kannan\nMicrosoft Research
\n\nTime: 4:00 - 4: 45 PM\n\nVersatility of Singular Value Decomposition\n
Singular Value Decomposition (SVD) is a basic tool to deal with matrix dat
a and has traditionally been applied in a variety of fields. In the modern
setting\, matrix sizes have increased\, but improved sampling based algor
ithms are still effective. Besides\, many new applications of SVD to Gauss
ian Mixtures\, Nearest Neighbors\, Topic Modeling etc. have been developed
. Combined with a simple device of thresholding\, SVD is useful on a new b
unch of problems. The talk will discuss from first principles some recent
developments.\n\n\nThese lectures jointly organized by ICTS and the Depart
ment of Computer Science and Automation\, IISc are a part of the Symposium
on Learning\, Algorithms and Complexity\n\nhttps://indico.tifr.res.in/ind
ico/conferenceDisplay.py?confId=4130
LOCATION:IISc campus\, Bangalore Auditorium\, Biology Department
URL:https://indico.tifr.res.in/indico/conferenceDisplay.py?confId=4130
END:VEVENT
END:VCALENDAR