Honours and Masters project

The ability to forecast human trajectory and/or body motion (i.e. pose dynamics and trajectory) from camera or other visual sensors is an essential component for many real-world applications, including robotics, healthcare, detection of perilous behavioural patterns in surveillance systems.

Visually discriminating the identity of multiple (similar looking) objects in a scene and creating individual tracks of their movements over time, namely multi-object tracking (MOT), is one of the basic yet most crucial vision tasks, imperative to tackle many real-world problems in surveillance, robotics/autonomous driving, health and biology.

A lot of decision support systems have been developed to predict or suggest a diagnosis about the health conditions of patients with the aim to assist clinicians in their decisional process. One of the techniques that is proved to present an efficient tool for medical healthcare decision making is Bayesian networks (BNs). BNs are recognized as efficient graphical models that can be used to explain the relationships between variables.

This multidisciplinary project combines cutting-edge Natural Language Processing (NLP), Chinese Studies and Political Science. The project aims to develop a deeper understanding of how official discourse has developed throughout the history of the People’s Republic of China. The main focus will be on text in the People’s Daily, the largest newspaper in China and the official newspaper of the Chinese Communist Party.

The brains of all bilaterally symmetric animals on Earth are divided into left and right hemispheres. The anatomy and functionality of the hemispheres have a large degree of overlap, but they specialize to possess different attributes. This principle is poorly understood and has not been exploited in AI/ML. Previously, we mimicked biological differences between hemispheres, and achieved specialization and superior performance in a classification task that matched behavioral observations.

The hippocampus is critical for episodic memory, a key component of intelligence, and a sense of self. There are a number of computational models, but none of them consider the fact that the hippocampus is, like the rest of the brain, divided into Left and Right hemispheres. Division into Left and Right is poorly understood, but undoubtedly critical, as it is a remarkably conserved feature of all bilaterally symmetric animals on Earth.

The brains of all bilaterally symmetric animals on Earth are divided into left and right hemispheres. The anatomy and functionality of the hemispheres have a large degree of overlap, but they specialize to possess different attributes. This principle is poorly understood and has not been exploited in AI/ML. The right hemisphere is more dominant for novelty, and the left for routine. Activity slowly moves to the left hemisphere as a task is perfected. In this project, we apply that principle to continual RL, where new tasks are introduced over time.

This project takes a different approach to RL, inspired by evidence that Hippocampus replays to the frontal cortex directly. It is likely used for model building, as opposed to the mainstream view in cognitive science and ML - where 'experience replay' ultimately improves policy. The predicted benefits are sample efficiency, better ability to generalize to new tasks and an ability to learn new tasks without forgetting old ones. The project objective is to improve biological models and advance state-of-the-art in continual reinforcement learning.

Given a knowledge base describing the existing background constraints and assumptions about what is possible in the world as well as the prior experience of an autonomous agent on the one hand and probabilistic perception of the current state of the world of the autonomous agent, on the other hand, it is essential to devise and efficiently enumerate the most consistent world models that are likely to be valid under the prior knowledge in order to refine the agent’s up-to-date perception and take the most suitable actions.

Mini-CP https://www.info.ucl.ac.be/~pschaus/minicp.html is a minimal form of constraint programming solver, designed to allow for easy experimentation and learning. 

One of the most efficient approaches to discrete optimisation solving is using lazy clause generation, which is a hybrid SAT/CP approach to solving problems.  But MiniCP does not currently support this.