Abstract
The remarkable success of deep learning is largely attributed to the collection of large datasets with human-annotated labels. However, it is extremely expensive and time-consuming to label extensive data with high-quality annotations. In other words, noisy samples are inevitable in large datasets. In this thesis, we aim to tackle the issues caused by noisy training samples under two noise-sensitive practical settings: instance recognition task and few-shot classification task.
Three contributions are made in this thesis. First, in chapter 3 we investigate the negative effect caused by noisy training samples in instance recognition task, which has been largely neglected by existing works. Specifically, we focus on person re-identification (re-ID)---a cross-domain instance matching problem---and propose to model uncertainty for features of input samples. This extra dimension allows the model to focus more on the clean inliers rather than overfitting to noisy training samples, resulting in better class separability and better generalisation to test data. Second, inspired by the ability of modelling uncertainty, in chapter 4 we focus on making full use of modelling uncertainty by encouraging neuron variance to build a unified framework for multiple applications, including network pruning, adversarial defence, and model calibration. Finally, we move on to few-shot classification problem in chapter 5. Since simultaneously optimising for high per-activation variability/uncertainty and predictive accuracy used in chapter 4 improves the few-shot learning model marginally, we propose hybrid graph neural networks to respectively overcome noisy training samples and class overlapping issues.