<div dir="ltr">A reminder about this exciting event!<div><br></div><div>See yinz there.</div><div><br></div><div>Artur<br><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">---------- Forwarded message ---------<br>From: <strong class="gmail_sendername" dir="auto">Suzanne Muth</strong> <span dir="auto"><<a href="mailto:lyonsmuth@cmu.edu">lyonsmuth@cmu.edu</a>></span><br>Date: Thu, Nov 2, 2023 at 11:42 AM<br>Subject: RI Ph.D. Thesis Proposal: Jack Good<br>To: RI People <<a href="mailto:ri-people@andrew.cmu.edu">ri-people@andrew.cmu.edu</a>><br></div><br><br><div dir="ltr"><div><div><p style="margin:0in;background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial"><span style="color:black"><font face="arial, sans-serif">Date: 14 November<span> 2023</span><br>
Time: 4:30 p.m. (ET)<br>
Location: GHC 8102</font></span></p><p style="margin:0in;background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial"><font face="arial, sans-serif"><span style="color:black">Zoom Link: </span><span style="color:black"><a href="https://cmu.zoom.us/j/96881722005?pwd=NjhCb2gwQ3ZzSmtlVlJ3Qnp5QTd1Zz09" style="color:blue" target="_blank"><span style="color:rgb(17,85,204);background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial">https://cmu.zoom.us/j/96881722005?pwd=NjhCb2gwQ3ZzSmtlVlJ3Qnp5QTd1Zz09</span></a></span><span style="color:black"> </span></font></p><p style="margin:0in;background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial"><font face="arial, sans-serif"><span style="color:black">Type: Ph.D. Thesis Proposal<br>
Who: Jack Good<br>
Title: </span><span style="background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial">Trustworthy Learning using Uncertain Interpretation of Data</span><span style="color:black"></span></font></p><p style="margin:0in;background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial"><span style="color:black"><font face="arial, sans-serif"> </font></span></p><p style="margin:0in;background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial"><span style="color:black"><font face="arial, sans-serif">Abstract:</font></span></p><p style="margin:0in;background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial"><font face="arial, sans-serif"><span style="background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial">Non-parametric models are popular in real-world
applications of machine learning. However, many modern ML methods that ensure
that models are pragmatic, safe, robust, fair, and otherwise trustworthy in
increasingly critical applications, assume parametric, differentiable models.
We show that, by interpreting data as locally uncertain, we can achieve many of
these without being limited to parametric or inherently differentiable models.
In particular, we focus on decision trees, which are popular for their good
performance on tabular data as well as ease of use, low design cost, low
computational requirements, fast inference, and interpretability. We propose a
new kind of fuzzy decision tree we call a kernel density decision tree (KDDT)
because the uncertain input interpretation is similar to kernel density
estimation.</span><br>
<br>
<span style="background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial">We organize the completed and proposed
contributions of this thesis into three pillars. The first pillar is robustness
and verification: we show improvement of robustness to various adverse
conditions and discuss verification of safety properties for FDTs and KDDTs.
The second pillar is interpretability: by leveraging the efficient fitting and
differentiability of our trees, we alternatingly optimize a parametric feature
transformation using gradient descent and the tree by refitting to obtain
compact, interpretable single-tree models with competitive performance. The
third pillar is pragmatic advancements: we make advances in semi-supervised
learning, federated learning, and ensemble merging for decision trees.</span></font></p><p style="margin:0in;background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial"><span style="color:black"><font face="arial, sans-serif"> </font></span></p><p style="margin:0in;background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial"><span style="color:black"><font face="arial, sans-serif">Thesis Committee Members:</font></span></p><p class="MsoNormal" style="background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial;margin:0in"><font face="arial, sans-serif">Artur Dubrawski, Chair</font></p><p class="MsoNormal" style="background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial;margin:0in"><font face="arial, sans-serif">Jeff Schneider</font></p><p class="MsoNormal" style="background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial;margin:0in"><font face="arial, sans-serif">Tom Mitchell</font></p><p style="margin:0in;color:rgb(0,0,0)">





















</p><p class="MsoNormal" style="background-image:initial;background-position:initial;background-size:initial;background-repeat:initial;background-origin:initial;background-clip:initial;margin:0in"><font face="arial, sans-serif">Gilles Clermont, University of Pittsburgh</font></p></div></div><div><div dir="ltr" class="gmail_signature" data-smartmail="gmail_signature"><div dir="ltr"><div dir="ltr" style="font-family:Calibri,Helvetica,sans-serif">
</div>
</div></div></div></div>
</div></div></div>