Dear all,
The talk by Samory Kpotufe is tomorrow:
Our seminar speaker this Friday in the thematic seminar is Samory
Kpotufe. Further below, there is also a list of upcoming talks that
are scheduled for the second semester.
Samory Kpotufe (Department of Statistics, Columbia
University, http://www.columbia.edu/~skk2175/)
Samory works on the intersection between statistics and machine
learning, with an interest in adaptive methods, and was one of the
chairs for last year's COLT learning theory conference.
Friday November 26, 16.00-17.00
Online on Zoom:
https://uva-live.zoom.us/j/85155421740
Meeting ID: 851 5542 1740
Please also join for online drinks after the talk.
Some Recent Insights on Transfer and Multitask Learning
A common situation in Machine Learning is one where training data is
not fully representative of a target population due to bias in the
sampling mechanism or due to prohibitive target sampling costs. In
such situations, we aim to ’transfer’ relevant information from the
training data (a.k.a. source data) to the target application. How
much information is in the source data about the target application?
Would some amount of target data improve transfer? These are all
practical questions that depend crucially on 'how far' the source
domain is from the target. However, how to properly measure
'distance' between source and target domains remains largely
unclear. In this talk we will argue that much of the traditional
notions of 'distance' (e.g. KL-divergence, extensions of TV such as
D_A discrepancy, density-ratios, Wasserstein distance) can yield an
over-pessimistic picture of transferability. Instead, we show that
some asymmetric notions of 'relatedness' between source and target
(which we simply term 'transfer-exponents') capture a continuum from
easy to hard transfer. Transfer-exponents uncover a rich set of
situations where transfer is possible even at fast rates; they
encode relative benefits of source and target samples, and have
interesting implications for related problems such as 'multi-task or
multi-source learning'. In particular, in the case of transfer from
multiple sources, we will discuss (if time permits) a strange
phenomena: no procedure can guarantee a rate better than that of
having a single data source, even in seemingly mild situations where
multiple sources are informative about the target. The talk is based
on earlier work with Guillaume Martinet, and ongoing work with Steve
Hanneke.
Seminar organizers:
Tim van Erven
Botond Szabo
https://mschauer.github.io/StructuresSeminar/
Upcoming talks:
Mar. 11, 2022, Tomer Koren, Tel Aviv University
Mar. 25, 2022, Nicolò Cesa-Bianchi, Università
degli Studi di Milano
Apr. 8, 2022, Julia
Olkhovskaya, Vrije Universiteit
Apr. 22, 2022, Tor Lattimore, DeepMind
--
Tim van Erven <tim@timvanerven.nl>
www.timvanerven.nl