Kyeong Won Lee Department of Statistics, Inha University

Bayesian Neural Networks Beyond the Curse of Dimensionality.

Neural networks have been widely used for complex and high-dimensional tasks, such as image analysis and natural language processing. One explanation for their success is that high-dimensional data often possess simpler underlying structures that neural networks can effectively exploit. In this seminar, we study this idea from a statistical perspective using Bayesian neural networks. In particular, we show how intrinsic low-dimensional structures can avoid the curse of dimensionality and lead to efficient estimation of complex functions. We also discuss how Bayesian neural networks can adapt to unknown function structures without prior knowledge of their complexity.