Self-adaptive Learning For Few-shot Image Classification

In recent years,deep learning methods have achieved great success in a wide range of applications.However,these models need large amounts of high-quality labeled data and many iterations to train the parameters.This requirement prevents the potential applications of the existing deep learning methods to the few-shot learning regime,in which only a few labeled examples are available for each category.We wonder whether the machines can acquire the ability to learn efficiently from only a few labeled samples.To tackle this problem,there has been a significant body of research work on few-shot image classification,which aims to learn a model that can classify unseen data instances(test samples)into a set of new categories,where only a few labeled instances(train samples)in each class are given.While promising,many existing methods suffer from two critical limitations.First,they do not take into account the feature importance of each instance and the feature correlation between train and test samples in each specific few-shot image classification task.Second,they are limited to supervised learning settings,as they require labeled data for training.However,unlabeled data are more abundantly available in real-world applications.To tackle the aforementioned challenges,we study self-adaptive learn-ing based methods for few-shot image classification.The main research contents and contributions are as follows:(1)Self-adaptive embedding for few-shot classification by hierarchical attention.The proposed method consists of two modules:the instance-aware attention module and the task-aware attention module,which can generate self-adaptive embeddings for train and test samples at multiple levels of granularity.First,the proposed instance-aware attention module can effectively capture the most discriminative features at the instance level,which can significantly improve the performance on downstream classi-fication tasks.Second,the proposed task-aware attention module can adaptively adjust the representations of train and test samples by considering the feature structures shared by them at the task level.Experiments on standard and cross-domain few-shot image classification tasks demonstrate that the proposed methods significantly outperforms several state-of-the-art methods.(2)Self-adaptive label augmentation for semi-supervised few-shot classification.The novelty of the proposed method lies in two folds.A major novelty of the proposed method is that for each learning task,our approach can effectively learn the task-adaptive metric in an end-to-end fashion by incorporating the intrinsic data property.Moreover,inspired by the observation that the pseudo-labels will be more reliable as the training course proceeds,we further propose a novel progressive neighbor selection strategy;that is,our approach only selects a small number of unlabeled samples with the most reliable pseudo labels for training and increases the number of them progressively.Experiments demonstrate that our proposed method outperforms several state-of-the-art methods on semi-supervised few-shot image classification tasks with or without distractors.