Decoupled Adaptive Convolutional Neural Networks for the Few-Shot Image Classification

Deep neural networks have achieved remarkable results in large-scale data domain. However, few-shot image classification is still a difficult and important problem. In this work, we analyze recently proposed deep neural networks of the meta-learning ability enable them to solve few-shot tasks, and try to find out where the ability comes from. Based on the analysis, we conclude a unified meta-learning framework for the few-shot tasks and propose a novel meta-learning network composed of two decoupled convolutional channels. One of them uses a typical convolutional neural network (CNN) to classify the query sample. The other channel is responsible for adaptively generating the parameters of the CNN with the support images. Unlike previous work in which the query and support input always share modules, our approach improves the flexibility and adaptability via decoupling the two channels. We demonstrate that our approach is able to achieve competitive performance with the existing state-of-the-art methods on two common few-shot image classification benchmarks while being substantially simpler and easier to train on.