This project explores how specialised programming frameworks called Guardrails, developed specifically for constraining large language models (LLMs), can prevent them from generating harmful, biased, or off-topic content. The goal of the project is to build simple examples using three leading guardrail frameworks implemented in Python: Guardrails AI, NeMo Guardrails from NVIDIA, and Llama Guard from … full description “A Comparative Analysis of Guardrail Frameworks for Large Language Models (available)”
Category: Artificial Intelligence
AI-powered COVID-19 detection via cough sounds (available)
Respiratory diseases such as COVID-19 are known to physically damage our airway and lungs, which in turn alters the produced respiratory sounds (e.g., cough, breadth). During the pandemic, cough classification has emerged as an accessible, low-cost, and environmentally friendly COVID-19 screening alternative, needing only a smartphone to collect and process cough samples. However, audio processing … full description “AI-powered COVID-19 detection via cough sounds (available)”
Algorithms for lifelong deep learning (available)
Continual learning [1], also known as lifelong learning, refers to the ability of an artificial intelligence system to continuously learn and adapt from new experiences over time. This is an important capability as it allows AI models to acquire new knowledge and skills as more data becomes available, without forgetting previously learned information. Continual learning … full description “Algorithms for lifelong deep learning (available)”
Building a Full Causality Chain Across an Enterprise System (completed)
Data Provenance refers to records of the inputs, entities, systems and process that influence data of interest, providing a historical record of the data and its origins. To provide a holistic view of the data provenance in an enterprise system, the provenance records of the activities carried out on a client workstation is important. Last … full description “Building a Full Causality Chain Across an Enterprise System (completed)”
Computer Vision for Extreme Environments (available)
The use of data from extreme environments in computer vision have shown an increase of interest in recent years as drones and autonomous vehicles were introduced into new uses. Nuclear plants, deep underwater and space vehicles are some of the areas computer vision can be applied to develop a fully autonomous system. Furthermore, the development … full description “Computer Vision for Extreme Environments (available)”
Cybercrime and ransomware groups – data analysis (available)
Project background Cybersecurity is the practice of defending computers, servers, mobile devices, electronic systems, networks, and data from malicious attacks, in an ever increasingly complex threat landscape. These attacks constitute a variety of computer-enabled and computer-dependent crimes, broadly categorised as ‘cybercrime’. In order to be in a position to defend against these attacks and minimise … full description “Cybercrime and ransomware groups – data analysis (available)”
Deep learning based underwater image segmentation and object recognition (available)
In the past decade, advances in marine object recognition have been dramatically boosted for monitoring of underwater ecosystems. Traditional statistical analysis and ocean model simulation heavily depend on the availability of visual features. However, Due to light attenuation and scattering problem, the underwater images captured by optical imaging system are heavily degraded. As a result, … full description “Deep learning based underwater image segmentation and object recognition (available)”
Deep Learning-based Environmental Sound Classification (available)
Automatic sound classification attracts increasing research attention owing to its vast applications, such as robot navigation, environmental sensing, musical instrument classification, medical diagnosis, and surveillance. Sound classification tasks involve the extraction of acoustic characteristics from the audio signals and the subsequent identification of different sound classes. In this project, we will explore diverse deep neural … full description “Deep Learning-based Environmental Sound Classification (available)”
Distilling large language models (available)
Model distillation for large language models (LLMs) presents a key challenge in AI research: how to compress massive, computationally expensive models into smaller, more efficient versions while preserving their performance. Large models, such as GPT-4, require vast amounts of memory and processing power, making them impractical for real-time applications on edge devices or personal computers. … full description “Distilling large language models (available)”
Embeded programming for INTI (available)
We are building a low-power transputer (parallel computer with microcontrollers) to study neuromorphic computing and run inference in low power devices (low power as in low computing power and low energy consumption). It is named the Incipient Neuromorphic Transputer Initiatiave in honour of Inti the Inca sun god. And because it is an incipient effort … full description “Embeded programming for INTI (available)”
Exploring optimization algorithms for deep sequence models (available)
Deep sequence models such as recurrent neural networks (RNNs) are a key type of architecture in modern deep learning, particularly for processing sequential data such as language text, speech, video, and time series data. RNNs have loops that allow information to persist and be passed from one step to the next, enabling them to effectively … full description “Exploring optimization algorithms for deep sequence models (available)”
Federated Machine Learning – Security and Privacy Evaluation Framework (available)
Project Description User data is essential for many of the modern business operations, especially related to the building consumer segmentation and profiling – for marketing and services personalisation. Dealing with user data has its positive and negatives, especially after the General Data Protection Regulation (GDPR) collecting and storing user’s personal data. Centralised machine learning approaches … full description “Federated Machine Learning – Security and Privacy Evaluation Framework (available)”
Implementation of biologically inspired efficient deep learning models (available)
As deep learning models continue to grow in size and complexity to tackle increasingly difficult tasks, the need for efficient and scalable models becomes ever more important. Extremely large language models like GPT-4 require massive computational resources and expensive hardware to train and run. This makes them impractical to deploy at scale in many real-world … full description “Implementation of biologically inspired efficient deep learning models (available)”
Machine Learning for Crystal Structure Prediction (completed)
Crystal Structure Prediction (CSP) is one of the major problems in computational chemistry with numerous applications in real life. This is essentially the (global) minimisation of a continuous, high-dimensional, complicated function. Many heuristic methods have been proposed for CSP and recently new methods based on Machine Learning were introduced. The goal of this project is … full description “Machine Learning for Crystal Structure Prediction (completed)”
Mixed Nash Equilibria in Net Coordination Games (completed)
Net Coordination Games form a special class of many-player games with several applications in Theoretical Computer Sciene, Multi Agent Systems, and Semi Supervised Learning. Nash equilibria correspond to the stable outcomes and they are the prominent solution concept in games. It is known that Net Coordination Games possess a pure Nash equilibrium, but unfortunately, it … full description “Mixed Nash Equilibria in Net Coordination Games (completed)”
Predicting and Explaining Drug Toxicity with Graph Neural Networks (available)
This student research project focuses on predicting and explaining drug toxicity using advanced machine learning techniques, specifically Graph Neural Networks (GNNs). Graph Neural Networks are a modern, powerful class of machine learning models widely used across diverse fields, from chemistry to social networks. However, one of the major challenges with traditional GNNs is their “black-box” … full description “Predicting and Explaining Drug Toxicity with Graph Neural Networks (available)”
Teaching Small LLMs to Reason (ongoing)
Large Language Models (LLMs) such as GPT4, are a game-changer for AI. Equipped with hundreds of billions of parameters, and trained on vast amounts of textual data totalling hundreds of terabytes, these models have revolutionised operations across numerous domains. But despite their considerable capabilities, their sheer size often means that they require substantial computational resources … full description “Teaching Small LLMs to Reason (ongoing)”
Video Action Classification (available)
Automatic interpretation of human actions from realistic videos attracts increasing research attention owing to its growing demand in real-world deployments such as biometrics, intelligent robotics, and surveillance. In this project, we will explore a variety of deep neural networks for video action classification, owing to their great efficiency in spatial-temporal feature learning.