Research

Our research is motivated by practical issues observed when applying deep learning models to real-world prediction problems, particularly in settings involving tabular and structured data. In many applications, models achieve strong performance on benchmarks but exhibit unstable behavior when deployed, such as sensitivity to spurious correlations or unexpected failures under distribution shifts.

Problems We Study

• Why do models trained on tabular data often rely on unstable or redundant features, even when overall prediction accuracy is high?
• How can representation collapse or loss of informative features be identified during training or after deployment?
• How can anomalous or out-of-distribution inputs be detected when labels are scarce or unavailable?

Our Perspective

• We use information-theoretic quantities as diagnostic tools to analyze the behavior of learned representations.
• Our focus is on lightweight methods that can be applied to existing models without extensive retraining or additional data.
• We emphasize analyses and algorithms that are feasible in practical deployment settings.

Application Contexts

• Tabular prediction problems in finance, healthcare, and industrial systems, where data are heterogeneous and often limited.
• Monitoring and anomaly detection scenarios in which reliability is as important as accuracy.