Would a GP forget her general commonsense knowledge (e.g., not to touch fire) to learn about human anatomy and diseases? As humans, we do not unlearn what we learned in order to learn a new skill. Shouldn't we expect the same from our very large language models (i.e., BERT ) as well?
Honours and Minor Thesis projects
Deep Generative Models such as Variational Autoencoders (VAEs)  have received a lot of attention in both text and image modelling. Their hierarchical flavour, Diffusion Models , has shown great generative capabilities in the text-to-image generation task. While successful, it is not clear how the learned representations via these models capture various semantic and syntactic aspects of the data.
IT Forensics is the art of extracting digital pieces of evidence also known as (aka) artifacts in a forensically sound manner, that is presentable to a court of law. In doing this it covers a range of conceptual levels, from high-level operating systems and computer theory down to computer networking.
The specific objective(s) of this project is to look at an encrypted piece of data and distinguish what encryption algorithm is used/employed. This would benefit IT Forensics researchers/investigators attacking encrypted volumes, files, folders, etc.
With advances in technology our energy networks are becoming modernised and smarter, enabling them to monitor and respond dynamically to local changes in energy demand. Creating effective and engaging visualisation methods can help with several aspects of improving efficiency, e.g. the control and distribution of electricity in operation, and raise awareness of energy conservation among consumers. One effective way would be leveraging Augmented Reality (AR), Virtual Reality (VR) or Mixed Reality (XR) to enable operators to visualize information in the context of physical infrastructure.
Medical guidelines provide human experts with steps on conducting diagnosis. While existing online knowledge graphs such as Unified Medical Language System (UMLS)  provide a wide coverage of various biomedical entities, their coverage is limited in certain specific domains such as Ophthalmology. In this project we will build an automated system that convert the offline medical guidelines of Glaucoma  and Diabetic Retinopathy  to machine readable knowledge graphs.
Biomedical Knowledge graphs (e.g., UMLS ) store structured relational information about medical and biological entities, e.g. entity-"Posttraumatic arteriovenous fistula" is connected to entity-"Traumatic arteriovenous fistula" via the relation "is associated morphology of". But these relations are very often too abstract or technical, and require human expertise to interpret or understand.
Knowledge graphs (KGs) play an important role in Natural Language Processing (NLP) and store information in a structured and machine-accessible way. They are used in different domains and fields. Generating texts from KGs is an important NLP task which transforms graph into natural language. For example, given a subgraph from a KG, we aim to get a corresponding description. Texts are easier to understand for human than graph-structured data.
Massive Open Online Courses (MOOCs), as one of the available options, are endowed with the mission to educate the world. MOOCs refer to online courses that are designed for an unlimited number of participants. In MOOCs, the learning materials are distributed over theWeb, which can be accessed by learners with internet connections anytime and anywhere. MOOCs are becoming increasingly popular. According to Class Central, by the end of 2022, there have been over hundreds of million learners enrolled in MOOCs in various MOOC platforms including edX, Coursera, etc.
Convolutional neural network (CNN) has exhibited its significance in addressing large-scale vision tasks such as action recognition, image classification, super-resolution and denoising. Recently, many researchers have reported the potential of employing deep learning techniques to refine high-quality clinical images for diagnosis and treatment.
Recently large-scale pre-trained language models, such as GPTs or BART have achieved successful performances in generating grammatically fluent text and capturing knowledge present in training corpus. However, it is pointed out that generating multi-sentence text (i.e., stories, narrations) with internal logical consistency is still far from being solved , with existing solutions merely scratching the surface in simple settings .