<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
</head>
<body>
<div class="" style="word-wrap:break-word; line-break:after-white-space">
<div class="" style="word-wrap:break-word; line-break:after-white-space">
<div dir="auto" class="" style="word-wrap:break-word; line-break:after-white-space">
<font face="Arial" class=""><b class="" style="font-size:14px">University of Chicago </b></font><b class="" style="font-family:Arial; font-size:14px">and Toyota Technological Institute at Chicago</b></div>
<b class="" style="font-family:Arial; font-size:14px">Machine Learning Seminar Series</b><br class="" style="font-family:Arial">
<br class="">
<div dir="auto" class="" style="word-wrap:break-word; line-break:after-white-space">
<font face="Arial" class=""></font></div>
</div>
<div class=""></div>
</div>
<div><img src="cid:2b8fdd4f-c01e-48a5-a34b-53279b7e0c07@namprd11.prod.outlook.com">
</div>
<div style="word-wrap:break-word; line-break:after-white-space">
<meta content="text/html; charset=utf-8">
<div class=""></div>
<div class="" style="word-wrap:break-word; line-break:after-white-space">
<div dir="auto" class="" style="word-wrap:break-word; line-break:after-white-space">
<font face="Arial" class=""></font></div>
<div dir="auto" class="" style="word-wrap:break-word; line-break:after-white-space">
<font face="Arial" class=""><br class="">
<b class="" style="font-size:14px">Samory Kpotufe, Associate Professor</b></font></div>
<div dir="auto" class="" style="word-wrap:break-word; line-break:after-white-space">
<font face="Arial" class=""><span class="" style="font-size:14px"><b class="">Columbia University</b></span></font></div>
<div dir="auto" class="" style="word-wrap:break-word; line-break:after-white-space">
<font class=""><br class="">
<font color="#000000" class="" style="font-family:Arial"><span class="" style="font-size:14px"><b class="">Friday, March 13, 10:30 - 11:30 am<br class="">
</b></span></font><font face="Arial" class="">JCL Rm 390</font></font></div>
<div dir="auto" class="" style="word-wrap:break-word; line-break:after-white-space">
<font class=""><font color="#000000" class=""><span class="">(Remote broadcast at TTIC Rm 526)</span></font><br class="">
<br class="">
<b class="" style="font-family:Arial; font-size:14px">Title:</b><b class=""><font face="Arial" class="" style="font-size:14px"> </font></b></font><font face="Arial" class="" style="font-size:14px">Some Recent Insights on Transfer Learning </font></div>
<div dir="auto" class="" style="word-wrap:break-word; line-break:after-white-space">
<font face="Arial Narrow, sans-serif" size="3" class=""><font color="#000000" class=""><br class="">
</font></font></div>
<div dir="auto" class="" style="word-wrap:break-word; line-break:after-white-space">
<font face="Arial" class=""><span class=""><font face="Arial" class="" style="font-size:14px"><b class="">Abstract:</b></font></span></font></div>
<div dir="auto" class="" style="word-wrap:break-word; line-break:after-white-space">
<font face="Arial" class=""><font class=""><span class=""><span class="" style="color:rgb(32,31,31); background-color:white; text-align:justify; font-size:14px">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 high costs in sampling the target population; 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? How much target data should we collect if any? 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. </span></span></font><font class=""><span class=""><span class="" style="color:rgb(32,31,31); background-color:white; text-align:justify; font-size:14px">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 new notions of ‘relative dimension’ 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, helps answer questions such as the benefit of unlabeled or labeled target data, yields a sense of optimal vs
suboptimal transfer heuristics, and have interesting implications for related problems such as multi-task learning. </span></span></font><span class="" style="color:rgb(32,31,31); background-color:white; text-align:justify; font-size:14px">Finally, transfer-exponents
provide guidance as to *how* to efficiently sample target data so as to guarantee improvement over source data alone. We illustrate these new insights through various simulations on controlled data, and on the popular CIFAR-10 image dataset. </span><span class="" style="color:rgb(32,31,31); background-color:white; text-align:justify; font-size:14px">The
talk is based on work with Guillaume Martinet, and ongoing work with Steve Hanneke.</span></font></div>
<div dir="auto" class="" style="word-wrap:break-word; line-break:after-white-space">
<font face="Arial" class=""><span class="">
<div class=""><font color="#000000" class=""><span class="" style="font-size:14px"><br class="">
</span></font></div>
<div class=""><font color="#000000" class=""><span class="" style="font-size:14px"><b class=""><br class="">
</b></span></font></div>
<div class=""><font color="#000000" class=""><span class="" style="font-size:14px"><b class="">Bio:</b></span></font></div>
<font color="#000000" class=""><span class="" style="font-size:14px">I graduated (Sept 2010) in Computer Science at the University of California, San Diego, advised by Sanjoy Dasgupta. I then was a researcher at the Max Planck Institute for Intelligent Systems.
At the MPI I worked in the department of Bernhard Schoelkopf, in the learning theory group of Ulrike von Luxburg. Following this, I spent a couple years as an Assistant Research Professor at the Toyota Technological Institute at Chicago. I then spent 4 years
at ORFE, Princeton University as an Assistant Professor. </span></font></span></font></div>
<div dir="auto" class="" style="word-wrap:break-word; line-break:after-white-space">
<font face="Arial" class=""><span class=""><font color="#000000" class=""><span class="" style="font-size:14px"> <br class="">
</span></font>
<div class=""><font class="" color="#000000"><span class=""><span class=""><b class="" style="font-size:14px">Host: Eric Jonas</b></span></span></font></div>
<div class=""><font class="" color="#000000"><span class=""><span class=""><b class="" style="font-size:14px"><br class="">
</b></span></span></font></div>
<div class=""><font class="" color="#000000"><span class=""><span class=""><b class="" style="font-size:14px">PDF:</b></span></span></font></div>
</span></font></div>
</div>
</div>
<div class="" style="word-wrap:break-word; line-break:after-white-space">
<div class=""><br class="">
</div>
<div class="">Tricia Baclawski<br class="">
Project Assistant IV<br class="">
Computer Science Department<br class="">
5730 S. Ellis - Room 212<br class="">
Chicago, IL 60637<br class="">
<a href="mailto:pbaclawski@uchicago.edu" class="">pbaclawski@uchicago.edu</a><br class="">
(773) 702-6854</div>
<br class="">
</div>
</body>
</html>