Boosting the Generalization Capability in Cross-Domain Few-shot Learning via Supervised Autoencoder
原文链接 https://openaccess.thecvf.com/content/ICCV2021/papers/Liang_Boosting_the_Generalization_Capability_in_Cross-Domain_Few-Shot_Learning_via_Noise-Enhanced_ICCV_2021_paper.pdf
Motivation
Different from general few-shot learning (FSL) where large-scale source dataset and few-shot novel dataset are from the same domain, target dataset and source dataset under Cross-domain few-shot learning (CDFSL) setting come from different domains, i.e. the marginal distributions of features of images in two domains are quite different.
Contributions
- This work is the first work proposes to use supervised autoencoder framework
to boost the model generalization capability under few-shot learning settings
- This work take reconstructed images from autoencoder as noisy inputs and let the model further predict their labels, which proves to further enhance the model generalization capability.
- The two-step fine-tuning procedure that does reconstruction in novel classes better adapts model to the target domain
Methodology (原文图片，侵删)
Problem Formulation
given: 1) source domain 
; 2)target domain 
there's a domain shift between them, the model is pretrained on and fine-tunes on
Each “N-way K-shot” classification task in target domain contains a support dataset 
and a query dataset 
. The support set contains N classes with K labeled images in each class and the query set contains images from the same N classes with Q unlabeled images in each class
Objective: CDFSL wants to achieve a high classification accuracy on the query set when K is small
文章基于transfer learning，先利用source domain dataset pretrain一个model， 同时为了使得该model有较强的泛化能力，提出了noise-enhanced SAE (NSAE)：NSAE not only
predicts the class labels of the inputs but also predicts the labels of the “noisy” reconstructions.
Pre-train on the source domain
对于来自 的sample， NSAE训练过程中不仅需要计算reconstruction loss， 还需要考虑reconstruction 出来的images重新输入回classification module得到的loss：
Fine-tune on the target domain
2 steps 微调：
1） In the first step, 我们在target domain 中抽取 support images 微调 autoencoder architecture， the model aims at minimizing reconstruction loss
2) In the second step, only the encoder is used to fine-tune on with the classification loss
Choices of loss functions
文章中pre-train 和 fine-tune使用的是不同的loss function，具体的选择可以参考原文