Visual understanding is a fundamental research topic in artificial intelligence, of which the goal is to enable machines to “describe what they saw”. It has been the key technique in many real-world applications, like building smart glasses to describe surrounding scenes to aid visually impaired people, constructing autonomous vehicles with visual sensors (like Tesla), and developing sensitive content detectors for video/image sharing website to protect teenagers. In the last decade, research for visual understanding has become even more prevalent due to the great success of deep learning, especially deep convolutional neural networks (DCNN). By feeding high-quality annotated training data into a fully supervised learning (FSL) engine, DCNN models could even surpass human-level performance in many visual understanding tasks, such as object classification and face recognition.
However, conducting FSL in real-world scenarios is challenging due to 1) there are potentially unlimited object categories in real life such that it is almost impossible to collect hundreds of labelled samples for each category; 2) there are ambiguity and difficulty in label annotation such that the training data contains many false labels which mislead visual understanding models; and 3) there are new categories emerging continuously such that a system must dynamically update itself with new data to address the catastrophic forgetting problem, which is hard for an FSL model. Indeed, to realise the full potential that emerging deep learning-fueled visual understanding systems can offer, we must make training of deep models with WEAK supervision.
Compared to FLS, WSL is still in its infancy with major breakthroughs having been made only in the last few years. More concretely, in this PhD project, we will study how to train object recognition and detection models from limited data amount. Then, we step further to the case where the label is noisy. On top of that, we investigate the continual learning problem in which the visual understanding system can dynamically learn to recognize new categories without forgetting previous ones. Doing so will allow the system to learn to recognize thousands of visual concepts from limited and low-quality training data in open environments, thus facilitating real-world applications. The fundamental goals are to develop novel weakly supervised learning algorithms for visual understanding.
Members of the Vision Graphics and Visualisation research group cover the full breadth of these timely and important developments and offer a diverse set of possible PhD projects in these areas. See https://www.aber.ac.uk/en/cs/study-with-us/pg/research-degrees/phd-vgv/ for more information.
Eligibility and Desired Student background
We are seeking an enthusiastic individual to join the Computer Science Department at Aberystwyth University, UK, with the following attributes:
- A minimum 2:1 undergraduate (BEng, MEng) and/or postgraduate masters’ qualification (MSc) in a science and technology field: Computer Science, Engineering, Mathematics, with specialisation in Computer Vision, Machine Learning and AI
- Appropriate IELTS score (overall score of 6.0 with no component below 5.5) or TOEFL
- Familiarity with machine learning and probabilistic models
- Relevant software knowledge and experience, for example Python and tensor frameworks (PyTorch or TensorFlow), C++, etc
- A driven, professional and independent work attitude
- Excellent written and verbal communication skills
- Excellent candidates can be recommended to AberDoc scholarship (https://www.aber.ac.uk/en/study-with-us/fees/postgrad/uk/research/aberdoc/, deadline: 14. Jan, 2022), which will cover their tuition fees (up to the UK rate of £4,500 per annum*) and a maintenance allowance of approximately £15,609 per annum* and access to a travel and conference fund (max. £500 per annum*) will also be provided.
- Outstanding international candidates can also be recommended to the Departmental Scholarship, which will reduce the international tuition fee to a home fee.
How to Apply
Applications through Aberystwyth’s electronic application process (How to apply: Study With Us , Aberystwyth University), include the following attachments in pdf form:
- Degree certificates and transcripts (if you are still an undergraduate, provide a transcript of results known to date)
- A statement no longer than 1000 words that outlines your research.
- Academic references - all scholarship applications require two supporting references to be submitted. Please ensure that your chosen referees are aware of the funding deadline (to be determined), as their references form a vital part of the evaluation process. Please include these with your scholarship application.