- All Categories
- Featured Events
- Alumni
- Application Deadline
- Arts
- Campus Discourse
- Careers
- BU Central
- Center for the Humanities
- Charity & Volunteering
- Kilachand Center
- Commencement
- Conferences & Workshops
- Diversity & Inclusion
- Examinations
- Food & Beverage
- Global
- Health & Wellbeing
- Keyword Initiative
- Lectures
- LAW Community
- LGBTQIA+
- Meetings
- Orientation
- Other Events
- Religious Services & Activities
- Special Interest to Women
- Sports & Recreation
- Social Events
- Study Abroad
- Weeks of Welcome
- MechE Seminar Series: Zuankai Wang 11:00 am
- Right to the City 11:00 am
- Aerial Silks Skills 114:00 pm
- Aerial Silks Skills 135:00 pm
- Aerial Silks Skills 155:15 pm
- Sunset Spin5:30 pm
- Ballet, Beginner 016:00 pm
- Barre Pilates Fusion6:00 pm
- Aerial Silks Intermediate 036:15 pm
- Aerial Silks Skills 177:30 pm
- SE PhD Final Defense: Nguyen Nguyen2:30 pm
SE PhD Final Defense: Nguyen Nguyen
MSE PhD Final Defense: Nguyen Nguyen
TITLE: Robust Model Selection for Interpretable Discovery of Latent Processes Under Misspecification
ADVISOR: Ioannis Paschalidis SE, ECE, BME
CHAIR: TBD
COMMITTEE: Jonathan Huggins (Mathematics), Alex Olshevsky (SE, ECE, CS), Brian Kulis (SE, ECE, CS)
ABSTRACT: The adoption of machine learning in scientific discovery is hindered by a significant gap: while models often excel at prediction, they may fail to identify latent structures that are interpretable, parsimonious, and robust under real-world model misspecification. This limitation is especially important in latent variable modeling, where the number of underlying processes, states, or factors must be selected from imperfect data and approximate model families. Likelihood-based criteria such as the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) tend to overfit in these conditions. And entropy-penalized criteria such as the Integrated Completed Likelihood (ICL), while more conservative, can still behave inconsistently across different settings. This dissertation addresses this challenge by developing and extending the Accumulated Cutoff Discrepancy Criterion (ACDC), a robust model selection framework for interpretable latent process discovery under misspecification. ACDC shifts the focus from predictive accuracy toward component-level adequacy by selecting the number of latent processes according to whether fitted components are sufficiently close to the data-generating structure up to a specified tolerance. This work presents three primary contributions: (1) the development of the general ACDC framework for robust selection of latent components in misspecified models, (2) an extension of ACDC to state-number selection in Hidden Markov Models, with applications to animal movement and biologging data, and (3) preliminary results of ACDC for selecting latent factors in masked variational autoencoders for causal representation learning. Together, these contributions establish ACDC as a flexible framework for robust latent process selection and demonstrate its promise across a variety of scientific applications.
| When | 2:30 pm - 5:30 pm on 18 June 2026 |
|---|---|
| Building | CDS 1101 |