Multi-task classification network for few-shot learning

Zhong Ji, Yuanheng Liu, Xuan Wang, Jingren Liu, Jiale Cao, YunLong Yu

Abstract: Semantic information provides both internal coherence within categories and distinctiveness between categories that surpass mere visual concepts. Semantic information has been employed in Few-Shot Learning (FSL) to achieve additional performance improvements. Previous methods usually combine support image and semantic information to classify query image. However, in FSL, it is challenging to train a model on a limited base dataset such that the model can effectively fuse or interact with both modalities and obtain better feature representation on the novel dataset. To address this problem, we propose a Multi-task Classification Network (MCN) to decompose the current classification problem into a image-image classification problem and a semantic-image classification problem. Considering the issue that the results of image-image classification and semantic-image classification may not always be trustworthy, we introduce an Uncertainty-Aware Decision Module (UADM) which biases the final classification result towards the result with lower uncertainty in the two types of classification. Extensive experimental results on three datasets have consistently shown that our proposed method achieves impressive results. Particularly, compared to the baseline, we achieved a 2–3% improvement on the CUB, SUN, and Flower datasets in both the 5-way 1-shot and 5-way 5-shot settings.