Source code of the method LWAU shown in "Layer-Wise Adaptive Updating for Few-Shot Image Classification"
URL: https://arxiv.org/abs/2007.08129
LWAU is evaluated with two backbones:Conv-4 and ResNet12
The performances on MiniImagenet is shown as the following Table.
The upper part of the table shows the meta-learning based few-shot learning methods with Conv-4 backbone.
The lower part shows the methods with ResNet12 backbone.
Except for improving the meta-learner's few-shot learning performance, LWAU can also greatly speed up the meta-learner's learning on the support set.
This is because when learning on novel few-shot learning tasks, frezing LWAU meta-learner's bottom layers will not damage the meta-learner's performance.
The comparison between LWAU and MAML when their bottom layers are frozen is shown in the following Figure.
At last, LWAU extracts sparser image representations.
If LWAU is helpful to your work, please cite our paper. Thanks!
@article{qin2020layer,
title={Layer-Wise Adaptive Updating for Few-Shot Image Classification},
author={Qin, Yunxiao and Zhang, Weiguo and Wang, Zezheng and Zhao, Chenxu and Shi, Jingping},
journal={IEEE Signal Processing Letters},
year={2020}
}