Title: Visual learning from imperfect data via inductive bias modelling
Abstract: Deep learning approaches have led to remarkable progress in visual learning. A key driving factor for this progress is the abundance of high-quality labelled data for use as training data. However, due to privacy, safety, and/or ethical issues, the availability and/or quality of such large-scale resources cannot be guaranteed in many tasks, which greatly hinders the rapid deployment of visual learning systems in the real world. Over the years, much effort has been directed towards seeking economic solutions from a variety of perspectives, yet the resultant models are far behind their supervised counterparts in terms of accuracy and performance.This thesis proposes alternative methods of supervised learning with inductive bias modelling, which do not rely on either exhaustive supervision or a balanced distribution of data. The core idea is to leverage the learned inductive biases to enhance the capacity of model generalisation based on inexact, incomplete, and imbalanced data. These inductive biases take various forms including neighbourhood affinity, historical experience, cross-domain knowledge and intrinsic reward. This thesis mainly focuses on four broad types of data imperfections in the respective vision tasks. The contributions of this thesis are (i) to enrich coarsely annotated data for semantic segmentation, where the given label information is propagated to unlabelled pixels with nonlocal graph modelling; (ii) to propose a novel concept of continual meta-learning for few-shot image classification, where the intra-taskand inter-task correlations can be well-preserved via message-passing and historical transition; (iii) to learn domain-agnostic concepts from unlabeled video data, where the representations of similar source and target videos are well mixed with bipartite graph neural networks; and (iv) to address the long-tail distribution issue for diverse visual paragraph generation, in which the model is incentivised to memorise precise yet rarely seen descriptions to context. Finally, we conclude by discussing open questions and future directions in imperfect data learning, with the aim to identify the key shortcomings and limiting assumptions of our existing approaches.
Publication Year: 2021
Publication Date: 2021-09-17
Language: en
Type: dissertation
Indexed In: ['crossref']
Access and Citation
AI Researcher Chatbot
Get quick answers to your questions about the article from our AI researcher chatbot