SFG: Sample-Centric Feature Generation for Semi-Supervised Few-Shot Learning
Published in IEEE Transactions on Image Processing, 2021
Abstract
Semi-supervised few-shot learning aims to improve the model generalization ability by means of both limited labeled data and widely-available unlabeled data. Previous works attempt to model the relations between the few-shot labeled data and extra unlabeled data, by performing a label propagation or pseudo-labeling process using an episodic training strategy. However, the feature distribution represented by the pseudo-labeled data itself is coarse-grained, meaning that there might be a large distribution gap between the pseudo-labeled data and the real query data. To this end, we propose a sample-centric feature generation (SFG) approach for semi-supervised few-shot image classification. Specifically, the few-shot labeled samples from different classes are initially trained to predict pseudo-labels for the potential unlabeled samples. Next, a semi-supervised metagenerator is utilized to produce derivative features centering around each pseudo-labeled sample, enriching the intra-class feature diversity. Meanwhile, the sample-centric generation constrains the generated features to be compact and close to the pseudo-labeled sample, ensuring the inter-class feature discriminability. Further, a reliability assessment (RA) metric is developed to weaken the influence of generated outliers on model learning. Extensive experiments validate the effectiveness of the proposed feature generation approach on challenging one- and few-shot image classification benchmark
Motivation
The motivation of the proposed sample-centric feature generation (SFG) on a 2-way-1-shot semi-supervised few-shot learning (SSFSL) task. Different colors denote features from different classes and the transparency represents the reliability of the augmented features (Less transparent means more reliable). Left: given only a single labeled sample for each class, there exist multiple decision boundaries (dashed lines). Middle: by pseudo-labeling features from the unlabeled set to augment the support set, the above challenge can be marginally alleviated. Right: by generating more diversified features surrounding each pseudo-labeled feature to augment the support set, the meta-learned decision boundary is more robust.
Framework
The framework of SPG is illustrated as follows.
Experimental Results
Our experiments are conducted on MiniImageNet, tieredImageNet, and CUB-100. The main experimental results:
Conclusion
In this work, we aim to fully leverage unlabeled samples that are readily available in real world, and have proposed a sample-centric feature generation (SFG) approach for semi-supervised few-shot learning. A sample-centric metageneration module is first designed to meta-learn from each pseudo-labeled sample for generating diversified yet discriminative new features. Then a RA metric is developed to weaken the negative impact of the generated feature outliers on model learning. Empirical studies on four benchmarks demonstrate that the proposed SFG greatly increases the classification accuracies of existing meta-learning models, especially for one-shot learning scenario where training data is very scarce.
When some of the unlabeled samples come from some distractor classes, the classification accuracy using such unlabeled samples will inevitably drop. Thus, how to generate diverse features (or samples) in the presence of the distractor classes needs to be further studied in the future.