Speaker：Professor Keng-Te Liao ( Institute of Statistics and Data Science, NTHU)
Topic：Representation Learning for Generalized Knowledge: From Identifying Invariance to Multimodal Fusion
Speaker：Professor Keng-Te Liao ( Institute of Statistics and Data Science, NTHU)
Time：Mar 27 (Friday) , 2026, 10:40-11:30
Place： 4F-427, Assembly Building I
Abstract
A fundamental challenge in deep learning lies in distinguishing true causal knowledge from transient correlations. While modern neural networks excel at high-dimensional pattern recognition, they often rely on features that appear predictive during training but fail to represent the underlying data-generating process. Consequently, these models often struggle with environment shifts, causing a lack of robustness when faced with out-of-distribution data.In this talk, I will present our research on learning and representing generalizable knowledge with deep neural networks. I will begin by introducing our work grounded in the invariance principle, where generalization is achieved by enforcing model behaviors remaining stable across diverse data-generating processes. Afterwards, I will extend these concepts to multi-modal fusion, addressing the challenge of learning unified latent representations from heterogeneous sources such as images, text, and audio. I will demonstrate the effectiveness of our work on identifying the underlying shared information, leading to robust performance and interpretable model behaviors in downstream tasks.