Matt's research aims to address the disconnect between state-of-the-art machine learning models, and models that are often used to solve real-world problems. Frequently, in real-world settings the modeller is confronted with a trade-off between maximising an objective (e.g., returning more accurate results) and minimising a budget (e.g., producing predictions in under a millisecond). His research considers three specific types of budgets: time, space, and privacy. These feature in recommendation, face recognition, bankruptcy prediction, stock market modelling, and real-time machine translation. Directly addressing these real-world trade-offs at an optimisation level results in algorithms that are simultaneously practical and accurate.