A Differentiable Architecture Search Approach for Few-Shot Image Classification

Abstract

AbstractFew-shot image classification is to learn models to distinguish between unseen categories, even though only a few labeled samples are involved in the training process. To alleviate the over-fitting problem caused by insufficient samples, researchers typically utilize artificially designed simple convolutional neural networks to extract features. However, the feature extraction capability of these networks is not strong enough to extract abstract semantic features, which will affect subsequent feature processing and significantly degrade performance when transferred to other datasets. This paper aims to design a general feature extraction network for few-shot image classification by improving the differentiable architecture search process. We propose a search space regularization method based on DropBlock and an early-stopping strategy based on pooling operation. Through the end-to-end search on the few-shot image dataset CUB, we obtain a light-weighted model FSLNet with excellent generalization ability. In addition, we propose a spatial pyramid self-attention mechanism to optimize the feature expression capability of FSLNet. Experiments show that the FSLNet searched in this paper achieves significant performance. The optimized FSLNet reaches state-of-the-art accuracy on the standard few-shot image classification datasets and in a cross-domain setting.KeywordsFew-shot image classificationDifferentiable architecture searchSpatial pyramid self-attention mechanismConvolutional neural networks