[106th TrustML Young Scientist Seminar] Talk by Wen-Bo Du (Nanjing University) "Non-Parametric Rehearsal Learning via Conditional Mean Embeddings"

Date and Time: October 8th, 2026, 10:30 -- 11:30 (JST)

Venue: Online

Title: Non-Parametric Rehearsal Learning via Conditional Mean Embeddings

Speaker: Wen-Bo Du (Nanjing University)

Abstract:
Machine learning is increasingly expected to go beyond prediction and guide actions. When an undesirable outcome is predicted, the next question is what actions can be taken to prevent it, a problem referred to as avoiding undesired future (AUF). Rehearsal learning has recently emerged as a new research direction for addressing AUF by leveraging influence relations. Existing methods, however, typically rely on restrictive parametric assumptions, limiting their applicability to complex real-world systems. In this talk, I will present our recent progress on a non-parametric rehearsal learning approach based on conditional mean embeddings. The proposed method formulates the AUF objective in a reproducing kernel Hilbert space, introduces a smooth Probit surrogate for the discontinuous desirability indicator, and estimates action-dependent outcome distributions through nested kernel ridge regression. The resulting estimator is identifiable from purely observational data and enjoys finite-sample error bounds and consistency guarantees. Experiments on synthetic and semi-synthetic benchmarks demonstrate the effectiveness and flexibility of the proposed approach.

Short Bio:
Wen-Bo Du is a Ph.D. student in the LAMDA Group at Nanjing University, supervised by Prof. Zhi-Hua Zhou. His research interests include machine learning, causal inference, rehearsal learning, and decision-making. His work has been published at leading conferences, including ICML and NeurIPS, with papers selected for oral and spotlight presentations. He also serves as a reviewer for conferences such as ICML, NeurIPS, and ICLR, as well as for the journal Machine Learning. He was selected as a Golden Reviewer for ICML 2026.