Abstracts

Novel meta-learning approaches for few-shot image classification

by Heda Song

In recent years, there has been rapid progress in computing performance and communication techniques, leading to a surging interest in artificial intelligence. Artificial intelligence aims to achieve human intelligence by making a machine think. However, current machine learning and optimisation techniques are far from fully accomplishing this, suffering from several limitations. For example, humans can learn a new concept quickly from very few examples, while artificial intelligence algorithms usually require a large number of examples to extract useful patterns. To tackle this issue, the computer science community has recently delved into the challenge of learning from very limited data, also known as few-shot learning. Few-shot image classification is the most studied research field of few-shot learning, which attempts to learn a new visual concept from limited labelled images. The conventional deep learning techniques cannot be simply applied to solve the problem, hindered by two core issues of few-shot learning, namely lack of information and intrinsic uncertainties. Lack of information is related to the insufficient visual patterns in limited training data, and intrinsic uncertainties are reflected by unrepresentative examples and background clutters. To tackle the problems, recent approaches mostly incorporate meta-learning methods which learn the general knowledge about how to make a few-shot learning process easier and quicker from a collection of learning tasks. However, existing meta-learning approaches mostly focus on either of the two key problems of few-shot image classification. Very few existing works consider both of them at the same time. Therefore, there is a need for developing novel meta-learning approaches that take into account both problems simultaneously for few-shot image classification. The thesis focuses on developing novel strategies of meta-learning approaches for few-shot image classification through three progressive stages, with the goal of addressing the aforementioned two core issues concurrently from different perspectives. In the first stage, we tackle the two main problems from the viewpoint of maximising the use of limited training data. Concretely, we propose learning to aggregate embeddings based on a channel-wise attention module. In this stage, we assume the embeddings after feature extraction consists of sufficient useful features. However, a feature extraction process could also lose relevant features. Hence, in the second stage, we target making sure as many useful features as possible can be extracted during a feature extraction process. Specifically, we design a spatial attention-based adaptive pooling module, in which a learnable pooling weight generation block is trained to assign different pooling weights to the features at different spatial locations. To further improve the classification performance, in the third stage, we leverage auxiliary information, such as saliency maps which can highlight the target object in an image, to compensate for the lack of…