<div dir="ltr">As previously announced, more intellectual cameos are coming our way in the first week of May.<div><br></div><div>Please mark your calendars for Cristian's thesis proposal talk:</div><div><br></div><div>Cheers</div><div>Artur</div><div><br><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">---------- Forwarded message ---------<br>From: <strong class="gmail_sendername" dir="auto">Diane Stidle</strong> <span dir="auto"><<a href="mailto:stidle@andrew.cmu.edu">stidle@andrew.cmu.edu</a>></span><br>Date: Wed, Apr 19, 2023 at 10:42 AM<br>Subject: Thesis Proposal - May 5, 2023 - Cristian Challu - Representing Time: Towards Pragmatic Multivariate Time Series Models<br>To: <a href="mailto:ml-seminar@cs.cmu.edu">ml-seminar@cs.cmu.edu</a> <<a href="mailto:ML-SEMINAR@cs.cmu.edu">ML-SEMINAR@cs.cmu.edu</a>>, <<a href="mailto:ywu@stat.ucla.edu">ywu@stat.ucla.edu</a>><br></div><br><br>
<div>
<p><i><b>Thesis Proposal</b></i></p>
<p>Date: May 5, 2023<br>
Time: 3:00pm (EST) (Remote only)<br>
</p>
<p><b>Title: </b><b>Representing Time: Towards Pragmatic
Multivariate Time Series Models</b></p>
<div dir="ltr">Abstract: </div>
<div dir="ltr">Time series models leverage temporal dependencies
among observations and interactions between multiple features in a
data stream. During the last decade, the unprecedented success of
deep learning (DL), primarily in its application to Computer
Vision and Natural Language Processing, has slowly but steadily
permeated to time series analytics. From Recurrent Neural Networks
to Transformers, new advancements in architectural design improved
capabilities and performance. Despite this success, I identify
several challenges to the broader adoption of current
state-of-the-art (SoTA) time series models, including distribution
shifts, missing data, computational complexity, trustworthiness,
and interpretability. The success of DL models is usually
justified by their ability to automatically discover helpful data
representations. We typically hope that multivariate time series
models, that involve high-dimensional objects with numerous time
series and temporal observations, exhibit useful temporal
dependencies and inter-feature relations. I propose to design new
DL architectures and algorithms for forecasting and anomaly
detection tasks that leverage these dependencies to induce
efficient learning of representations capable of (i) improving the
models' performance, (ii) improving robustness by favoring domain
adaptation, and (iii) reducing overparameterization to improve
scalability. The already completed work includes five different
models and algorithms that achieve SoTA performance on relevant
tasks and address a few key adoption challenges. The proposed
research extends that work in the following directions: developing
healthcare applications, analyzing the enabling conditions for
transfer learning between domains and tasks, and extending prior
work to multivariate long-horizon forecasting.</div>
<div dir="ltr"><br>
</div>
<div><b>Thesis Committee:</b></div>
<div>Artur Dubrawski, Chair</div>
<div>Roni Rosenfeld</div>
<div>Barnabas Poczos</div>
<div>Ying Nian Wu (UCLA)<br>
</div>
<div><br>
Link to the draft document: <a href="https://drive.google.com/drive/u/3/folders/1fyyY_h9Ok3zk1a5KzTgsJNCq4mIRhh9G" target="_blank">https://drive.google.com/drive/u/3/folders/1fyyY_h9Ok3zk1a5KzTgsJNCq4mIRhh9G</a><br>
<br>
Zoom meeting link: <a href="https://cmu.zoom.us/j/92381510152?pwd=ZXk0Q0JaNlVVSk55a3RXZkM1c2ZYQT09" target="_blank">https://cmu.zoom.us/j/92381510152?pwd=ZXk0Q0JaNlVVSk55a3RXZkM1c2ZYQT09</a><br>
</div>
<div><br>
</div>
<p><b></b></p>
<pre cols="72">--
Diane Stidle
PhD Program Manager
Machine Learning Department
Carnegie Mellon University
<a href="mailto:stidle@andrew.cmu.edu" target="_blank">stidle@andrew.cmu.edu</a></pre>
</div>
</div></div></div>