A Simple and Provable Method to Adapt Pre-trained Model across Domains with Few Samples

Danruo DENG, Guangyong Chen, Furui Liu, Jinpeng Li, Jianye HAO, Pheng-Ann Heng

Keywords: Cross-domain few-shot learning, PAC-Bayesian framework, dimensionality reduction, ensemble

TL;DR: This paper proposes a Simple and Provable method to quickly adapt a given pre-trained model across domains with few samples.

Abstract: Adapting the pre-trained model across domains with few samples, known as cross-domain few-shot learning, is a challenging task in statistical machine learning. Most previous efforts focused on training robust and transferable feature representations but rarely explored how to train an accurate few-shot model from a given pre-trained model. In this paper, we are interested in the performance of training a cross-domain few-shot classifier with representations from different layers of a pre-trained model and the impact of reducing the dimensionality of these representations. Based on this, we propose a simple and provable method, Average Pooling Ensemble Few-shot Learning (APEF). We demonstrate the effectiveness of average pooling and ensemble in cross-domain few-shot image classification both theoretically and experimentally. In particular, we provide a theoretical analysis in the PAC-Bayesian framework to illustrate why our method works, and we also empirically evaluate our approach on the challenging CD-FSL benchmark, which shows that our proposed method consistently outperforms all baselines.

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: Deep Learning and representational learning