Distribution Calibration For Few-Shot Learning by Bayesian Relation Inference

Yuchen Liu, Yang Gu, Weining Weng, Zhaohua Yang, Yiqiang Chen

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

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.

Keywords: Bayesian inference, few-shot learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We propose a distribution calibration method based on Bayesian relational inference for few-shot learning on medical Images and others.

Abstract: Learning from a limited number of samples is difficult as a small number of samples cannot cover all the information in their category. It is worth noting that categories with scarce samples may be distributed in a way that is related to categories that contain sufficient data. Therefore it is possible to calibrate the distribution of a sample-poor category by using categories with a large amount of data. Existing methods of distribution calibration usually use artificially set distances to calculate the association between two categories, which may ignore deeper relations between categories. In this paper, we propose a distribution calibration method based on Bayesian relation inference. For the input few-sample classes, it can automatically infer their relation with the categories with sufficient data and adaptively generate a large amount of fused feature data that can represent the few-sample classes. The results show that a simple logistic regression classifier trained by using the large amount of data generated by our method, exceeds state-of-the-art accuracy for skin disease classification issue. Through visual analysis, we demonstrate that the relationship graph generated by this Bayesian relationship inference method has a degree of interpretability.

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.

Submission Number: 6645