Honours and Masters project

Is the user paying attention? Is the content engaging enough?

The degree of concentration, enthusiasm, optimism, and passion displayed by individual(s) while interacting with a machine is referred to as ‘user engagement’. Engagement is a positive psychological state characterized by active behavioral participation, positive emotional experiences, and intense cognitive focus. Being able to detect engagement and/or attention has wide applications in consumer commerce, smart cars, augmented reality etc. 

Is the user paying attention? Is the content engaging enough?

The degree of concentration, enthusiasm, optimism, and passion displayed by individual(s) while interacting with a machine is referred to as ‘user engagement’. Engagement is a positive psychological state characterized by active behavioral participation, positive emotional experiences, and intense cognitive focus. Being able to detect engagement and/or attention has wide applications in consumer commerce, smart cars, augmented reality etc. 

Goal recognition is defined as the problem of determining an agent’s intent from observations of its behaviour. Current research in goal recognition has focused on observing agents that are trying to achieve their goals in a rational manner. Other research has focused on observing agents that are deliberately trying to trick an observer into believing they are pursuing alternative goals to the ones they are actually pursuing. However there is also a need to recognise when a behaviour is suspicious, regardless of the goal that is being tried to be achieved.

Markov Decision Processes (MDPs) are frameworks used to model decision-making in situations where outcomes are partly random and partly under the control of a decision maker. While small MDPs are inherently interpretable for people, MDPs with thousands of states are difficult to understand by humans. The K-MDP problem is the problem of finding the best MDP with, at most, K states by leveraging state abstraction approaches to aggregate states into sub-groups. The aim of this project is to measure and improve the interpretability of K-MDP approaches using state-of-the-art XAI approaches.

As intelligent agents make decisions, any project aiming to realize human-like AGI should model decision-making.  As we have been pursuing the WBA approach to create AGI by learning from the architecture of the entire brain, we request you to model the decision-making of the mammalian brain.

This project aims to develop privacy-preserving deepfake detection techniques that enable accurate and secure identification of synthetic audio and video content without exposing sensitive user data. Traditional detection methods often require access to raw audio or visual inputs, raising significant privacy concerns, especially in scenarios involving personal or biometric data.

This project focuses specifically on LLM applications: chatbots used in customer support (e.g., healthcare). The goal is to investigate how user requirements (e.g., “the bot should de-escalate frustrated users”) can be systematically translated into prompt templates or prompt strategies.

The research project aims to investigate:

- Multi-Model Fusion with Deep Neural Networks for Future Energy Systems (Smart Grid). 

Future energy systems are envisioned to be running decentrally with full automatic control, high proportion of renewable energy (e.g., wind & solar), and abundant storage facilities. With many types of renewable energy sources are weather and climate dependent, accurate and timely prediction on reliability risks (e.g., loss of generation, voltage issues, and thermal limit violations) due to weather/climate are often necessary.

Large language models (LLMs) have recently made significant progress in machine translation quality [1], but they still struggle with maintaining consistency and accuracy across entire documents. Professional translators commonly use translation memory (TM) tools to reuse past translations, ensuring consistent terminology and phrasing throughout a document.

This project will involve benchmarking state of the art methods for time series classification on the new MONSTER benchmark datasets [1, 2, 3].  Currently almost all benchmarking in time series classification is performed on the (almost all very small) datasets in the UCR and UEA archives.  This is particularly unsuitable for deep learning models which are low bias models and ideally trained using large quantities of data.  The "true" performance of current deep learning methods for time series classification is unknown outside of the UCR/UEA datasets.  Most deep learning models for times…