A1507
Title: Estimation error and hypothesis testing for non-identifiable models, with applications to machine learning
Authors: Junichiro Yoshida - The Institute of Statistical Mathematics (Japan) [presenting]
Abstract: To bridge machine learning and classical statistical methods, it is important to analyze estimation error in non-identifiable models such as neural networks. However, in complex non-identifiable models, estimation error is difficult to evaluate explicitly. To address this problem, an estimator of the estimation error is proposed together with its confidence interval, thereby enabling classical statistical methods, such as hypothesis testing, for non-identifiable models. The key idea is to apply the resolution of singularities to singular models following Watanabe. While his theory is primarily developed in a Bayesian framework, the estimator is feasible within a frequentist asymptotic framework and is computationally tractable.
