About the Project
Overview
Modern artificial intelligence (AI) increasingly relies on combining multiple sources of information, such as sound, motion, images, and sensor data, to achieve robust and intelligent behaviour. However, most existing multimodal AI systems depend on large, computationally intensive models that are difficult to deploy in settings with limited resources. This PhD project explores a different question: how can multimodal intelligence be achieved using minimal, efficient representations under resource constraints?
The project will investigate Tiny Multimodal Learning, a research direction focused on understanding what is essential for multimodal learning when computation, memory, or energy are limited. Rather than scaling up models, the research aims to identify principled, lightweight methods for representing and combining multimodal information over time. Grounded in machine learning, representation learning, and efficient algorithms, the work addresses real-world challenges in sustainable and accessible AI.
As a PhD researcher, you will gain experience in multimodal machine learning, experimental benchmarking, and algorithmic design, with opportunities to explore connections to dynamical systems and efficient model architectures. You will work in a supportive and inclusive doctoral environment, with access to strong research facilities, interdisciplinary collaboration, training in research and professional skills, and opportunities to publish at leading international venues.
Aims and objectives
The aim of this PhD project is to develop principled and efficient methods for multimodal learning under resource constraints. The research seeks to understand how different types of data—such as audio, motion, images and sensor signals—can be combined using minimal representations without relying on large, computationally intensive models.
Key objectives include:
Reviewing and benchmarking existing multimodal learning approaches under limited computational resources
Investigating lightweight representations and fusion strategies for multimodal data
Studying how temporal structure and missing data affect multimodal learning
Identifying fundamental trade-offs between accuracy, efficiency, and robustness.
The project emphasises foundational machine learning research while maintaining strong relevance to real-world, resource-aware AI systems.