Improve Novel Class Generalization By Adaptive Feature Distribution for Few-Shot Learning

Ran Tao, Marios Savvides

Keywords: Novel Class Generalization, Finetuning One Scale Vector, Adaptive Feature Distribution, Cross-Domain

Abstract: In this work, we focus on improving the novel class generalization of few-shot learning. By addressing the difference between feature distributions of base and novel classes, we propose the adaptive feature distribution method which is to finetune one scale vector using the support set of novel classes. The scale vector is applied on the normalized feature distribution and by using one scale vector to reshape the feature space manifold, we obtain consistent performance improvement for both in-domain and cross-domain evaluations. By simply finetuning one scale vector using 5 images, we observe a $2.23\%$ performance boost on 5-way 1-shot cross-domain evaluation with CUB over statistics results of 2000 episodes. This approach is simple yet effective. By just finetuning a single scale vector we provide a solution of reducing number of parameters while still obtain generalization ability for few-shot learning. We achieve the state-of-the-art performance on mini-Imagenet, tiered-Imagenet as well as cross-domain evaluation on CUB.

One-sentence Summary: By simply finetuning one scale vector on the normalized feature distribution which shows great novel class generalization for in-domain and cross-domain few-shot learning.

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

Reviewed Version (pdf): https://openreview.net/references/pdf?id=wlDv3te912