Speakers

Lars Kai Hansen

Professor, Technical University of Denmark (Denmark)

Missing-Data-Induced Phase Transition in Spectral PLS for Multimodal Learning

Partial Least Squares (PLS) learns shared structure from paired data via the top singular vectors of the empirical cross-covariance (PLS-SVD), but multimodal datasets often have missing entries in both views. We study PLS-SVD under independent entry-wise missing-completely-at-random masking in a simple high-dimensional model. We find theoretically and empirically that PLS-SVD exhibits a sharp BBP-type phase transition: below a critical signal-to-noise threshold the leading singular vectors are asymptotically uninformative, while above it they achieve nontrivial alignment with the "planted" shared directions. The theoretical analysis provides closed-form asymptotic overlap formulas.

Latha Pemula

Applied Scientist at AWS AILabs (USA)

Are we there yet?

Academia regularly celebrates zero-shot visual anomaly detection performance exceeding 90%. Yet, on the factory floor, an industrial revolution has failed to materialize. Real-world adoption remains stubbornly low because benchmark leaderboards mask severe deployment vulnerabilities. This keynote dissects the hidden barriers preventing production-ready deployment and outlines the modeling shifts needed for a true breakthrough.

Tomoya Nishida

Hitachi (Japan)

Anomalous Sound Detection: Lessons from Competitions and Beyond

Anomalous sound detection is an important technology for industrial monitoring. In real-world applications, however, its performance is often degraded by challenges such as lack of anomalous sounds, domain shifts, limited validation data, and environmental noise. The DCASE Challenge Task 2 has been held annually as a shared benchmark for addressing these issues. This talk reviews recent approaches to such challenges, highlighting the evolution of the benchmark and drawing on insights from the competition results, proposed methods, and related research. It also discusses important issues beyond anomaly detection itself.

Veronika Cheplygina

Professor, IT University of Copenhagen (Denmark)

Curious findings about medical image datasets

It may seem intuitive that we need high quality datasets to ensure for robust algorithms for medical image classification. With the introduction of openly available, larger datasets, it might seem that the problem has been solved. However, this is far from being the case, as it turns out that even these datasets suffer from issues like label noise and shortcuts or confounders. Furthermore, there are behaviours in our research community that threaten the validity of published findings. In this talk I will discuss both types of issues with examples from recent papers.