<div dir="auto"><div>Willa will be giving her proposal presentation on Monday next week during our usual brainstorming session time slot, but in Gates Hall. Please come and join. Intellectual fun guaranteed!</div><div dir="auto"><br></div><div dir="auto">Cheers,</div><div dir="auto">Artur<br><br><div class="gmail_quote" dir="auto"><div dir="ltr" class="gmail_attr">---------- Forwarded message ---------<br>From: <strong class="gmail_sendername" dir="auto">RI PhD Program Manager</strong> <span dir="auto"><<a href="mailto:ri-phd-manager@andrew.cmu.edu" target="_blank" rel="noreferrer">ri-phd-manager@andrew.cmu.edu</a>></span><br>Date: Mon, Apr 27, 2026, 1:52 PM<br>Subject: RI PhD Thesis Proposal - Willa Potosnak<br>To: RI People <<a href="mailto:ri-people@andrew.cmu.edu" target="_blank" rel="noreferrer">ri-people@andrew.cmu.edu</a>><br></div><br><br><div dir="ltr"><a href="https://www.ri.cmu.edu/event/forecasting-at-scale-with-efficient-deep-learning-architectures/" rel="noreferrer noreferrer" target="_blank"><b>RI EVENT CALENDAR</b></a><br><div><br></div><div><div id="m_7948080626998269783m_1322197882729342007gmail-:2w9" style="direction:ltr;margin:8px 0px 0px;padding:0px;font-size:0.875rem;overflow-x:hidden;font-family:"Google Sans",Roboto,RobotoDraft,Helvetica,Arial,sans-serif"><div id="m_7948080626998269783m_1322197882729342007gmail-:2w8" style="direction:ltr;font-style:normal;font-variant:normal;font-size-adjust:none;font-kerning:auto;font-feature-settings:normal;font-stretch:normal;font-size:small;line-height:1.5;font-family:Arial,Helvetica,sans-serif;overflow:auto hidden"><div id="m_7948080626998269783m_1322197882729342007gmail-avWBGd-2088"><div dir="ltr"><div><span style="color:inherit"><div><span style="font-family:arial,sans-serif"><b>Date:  </b>May 4th, 2026</span></div><div><span style="font-family:arial,sans-serif"><b>Time:  </b>4:30 PM-6:00 PM</span></div><div><span style="font-family:arial,sans-serif"><b>Location</b>: GHC 4405<br><a href="https://cmu.zoom.us/j/93920625206?pwd=WsTYphnRXRfhUaBpKKbbeIaTgDMoro.1" rel="noreferrer noreferrer" target="_blank"><b>Zoom Link</b></a></span></div><div><span style="font-family:arial,sans-serif"><b>Type: </b>RI PhD Thesis Proposal</span></div><div><span style="font-family:arial,sans-serif"><b>Who: </b>Willa Potosnak</span></div></span></div><span style="color:inherit"><div><span style="font-family:arial,sans-serif"><br></span></div><div><span style="font-family:arial,sans-serif"><b>Title:</b> Forecasting at Scale with Efficient Deep Learning Architectures</span></div></span><span style="color:inherit"><div><span style="font-family:arial,sans-serif"><br></span></div><div><b style="font-family:arial,sans-serif">Abstract:</b></div><div>Time Series Foundation Models (TSFMs) have scaled rapidly, with publicly reported pretraining corpora growing from 1.23 billion to 1 trillion data points between 2024 and 2026, an approximately 800× increase in two years. Recent work has further supplemented real-world data with synthetic data to expose models to broader time series patterns. Yet, this data-centric paradigm raises a fundamental question: <i>must intelligent forecasting rely solely on scale, or can intentional architectural design unlock better generalization? </i>This thesis proposes that more intelligently and efficiently leveraging existing data, rather than scale alone, is key to achieving better forecasting generalization. We pursue this through three parallel architectural themes: exploiting cross-channel structure beyond temporal patterns, enabling zero-shot generalization through structured composition, and reducing gradient and forecast variance by design. Each theme aims to enhance generalization with available data while treating computational efficiency as a core design principle.<br><br>In this thesis, we demonstrate that scale is not the only path to generalization by: developing multivariate architectures that leverage cross-channel dependencies efficiently while reducing forecast error; showing that architectures can generalize beyond their training distribution in both patterns and concepts; and verifying variance-aware architectural designs that extract richer training signals from existing data, provably reducing gradient variance while reducing forecast error and improving calibration.<br><br>Within the first theme, we further propose pretraining strategies for multivariate TSFMs to investigate whether data balancing and curriculum learning can improve downstream generalization given the same pretraining corpora. Within the second theme, we propose an additional dimension of generalization, extending beyond pattern and concept generalization to horizon generalization, an important consideration for TSFMs applied across diverse tasks and domains. Overall, this work contributes new insights into advancing time series forecasting generalization through efficient architectural design.</div><div><br></div><div><div><span style="font-family:arial,sans-serif"><b>Committee:</b></span></div><div>Artur Dubrawski, Chair<br>John Dolan<br>Barnabás Póczos<br>Michael W. Mahoney (University of California, Berkeley)</div></div><div><br></div><div><a href="https://drive.google.com/file/d/1oYYdgxvLaW4iF-LWHqzpWHMuXA677QYv/view?usp=sharing" style="font-family:arial,sans-serif" rel="noreferrer noreferrer" target="_blank"><b>Thesis Link</b></a></div></span></div></div></div></div><div id="m_7948080626998269783m_1322197882729342007gmail-:2vz" style="font-size:0.875rem;margin:15px 0px;clear:both;font-family:"Google Sans",Roboto,RobotoDraft,Helvetica,Arial,sans-serif"></div><br></div></div>
</div></div></div>