<div dir="auto"><div dir="auto">Hi all,</div><div dir="auto"><br></div><div dir="auto">I'll be proposing this Thursday at noon. Feel free to drop by.<br><br><div data-smartmail="gmail_signature" dir="auto">-kirthevasan<br><br>Sent from my phone. Please excuse brevity.</div></div><div class="gmail_quote">---------- Forwarded message ----------<br>From: "Diane Stidle" <<a href="mailto:diane%2B@cs.cmu.edu" target="_blank">diane+@cs.cmu.edu</a>><br>Date: Mar 27, 2017 14:17<br>Subject: Thesis Proposal - 4/6/17 - Kirthevasan Kandasamy - Tuning Hyper-parameters without Grad Students: Scaling up Bandit Optimisation<br>To: "<a href="mailto:ml-seminar@cs.cmu.edu" target="_blank">ml-seminar@cs.cmu.edu</a>" <<a href="mailto:ML-SEMINAR@cs.cmu.edu" target="_blank">ML-SEMINAR@cs.cmu.edu</a>>, "<a href="mailto:zoubin@eng.cam.ac.uk" target="_blank">zoubin@eng.cam.ac.uk</a>" <<a href="mailto:zoubin@eng.cam.ac.uk" target="_blank">zoubin@eng.cam.ac.uk</a>><br>Cc: <br><br type="attribution">
<div text="#000000" bgcolor="#FFFFFF">
<p><i>Thesis Proposal</i></p>
<p>Date: 4/6/17<br>
Time 12:00pm<br>
Place: 6121 GHC<br>
Speaker: Kirthevasan Kandasamy</p>
<p>Title: Tuning Hyper-parameters without Grad Students: Scaling up
Bandit Optimisation</p>
<div>Abstract:<br>
In many scientific and engineering applications, we are tasked
with optimising a black-box function which is expensive to
evaluate due to computational or economic reasons. In <i>bandit
optimisation</i>, we sequentially evaluate a noisy function with
the goal of identifying its optimum in as few evaluations as
possible. Some applications include tuning the hyper-parameters of
machine learning algorithms, on-line advertising, optimal policy
selection in robotics and maximum likelihood inference in
simulation based scientific models. Today, these problems face new
challenges due to increasingly expensive evaluations and the need
to perform these tasks in high dimensional spaces. At the same
time, there are new opportunities that have not been exploited
before. We may have the flexibility to approximate the expensive
function by investing less resources per evaluation. We can also
carry out several evaluations simultaneously, say via parallel
computing or by concurrently conducting multiple experiments in
the real world. In this thesis, we aim to tackle these and several
other challenges to meet emerging demands in large scale bandit
applications. We develop methods with theoretical underpinnings
and which also enjoy good empirical performance.<br>
<br>
Thesis Committee:<span><br>
<font size="-1" face="Helvetica, Arial, sans-serif">Barnabás
Póczos</font></span> (Co-Chair)<br>
Jeff Schneider (Co-Chair)<br>
Aarti Singh<br>
Zoubin Ghahramani (University of Cambridge)<br>
<br>
Link to draft document: <a href="http://cs.cmu.edu/%7Ekkandasa/docs/proposal.pdf" target="_blank">cs.cmu.edu/~kkandasa/docs/pro<wbr>posal.pdf</a></div>
<pre class="m_1841397363111358464m_8907962190129999609moz-signature" cols="72">--
Diane Stidle
Graduate Programs Manager
Machine Learning Department
Carnegie Mellon University
<a class="m_1841397363111358464m_8907962190129999609moz-txt-link-abbreviated" href="mailto:diane@cs.cmu.edu" target="_blank">diane@cs.cmu.edu</a>
<a href="tel:(412)%20268-1299" value="+14122681299" target="_blank">412-268-1299</a></pre>
</div>
</div></div>