Russel Dsouza's research focuses on making large language models more efficient, reliable, and understandable. At the Institute for Data and AI at the University of Birmingham, he contributes to this goal while also helping to shape research priorities and supporting data-intensive research proposals across the University.
Russel's individual research investigates the core challenges of model alignment and factuality. This includes analysing the sources of human-model disagreement in the data used to align large language models and contributing to the development of optimised fact-checking pipelines. This work is complemented by interdisciplinary collaborations. With colleagues in Psychology, he has explored the differences between human and model reasoning, leading to a system for automating the scoring of psychological assessments. With a team in Education, Russel's work involved analysing the impact of student-AI interactions.
Russel holds a Master of Science in Machine Learning from the University of Birmingham, where his dissertation focused on evaluating adversarial alignment subversion in large language models. He also holds a Bachelor of Technology in Electronics and Communications Engineering; his dissertation involved developing a deep learning system for the pixel-wise classification of land-use from satellite imagery. Russel's prior research experience includes developing machine learning models for applications in satellite imaging and computational histopathology.