BIL: Bandit Inference Learning for Online Representational Similarity Test

Xiaoting Ji, Shuoxun Xu, Wenhai Cui, Linglong Kong, Xiaodong Yan

Keywords: two-armed bandit process, online learning

TL;DR: Based on two-armed bandit process, this article proposes a strategic aggregation procedure for online representational similarity testing.

Abstract: Similarity analysis is commonly used to determine the size of the discrepancy between two representations of a distribution pattern. In contrast to classical representational similarity analysis, which identifies disparate types of representations based on their shared similarity structures in distance matrices, this article proposes an online hypothesis testing procedure that will be able to determine whether a representation's difference from a constant is more significant than a predefined margin for streaming data. As a basic reinforcement learning model, two-armed bandits (TAB) are used to construct test statistics that update online. To achieve the most efficient testing results, an optimal strategy is developed for the TAB process. Asymptotic test statistics are discussed in theory, as are its corresponding explicit density functions, which are more accumulated than the normal distribution commonly applied in classical statistical analysis. Since the power of the proposed representative similarity test (RST) method is higher than that of the classical test, simulation studies support the validity of the proposed method.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)