Antonio Grassi, Aristeidis Panos, University of Cambridge
Machine learning models are typically trained on fixed datasets, yet real-world applications often require them to adapt as new data, tasks, and environments emerge. Continual learning seeks to address this challenge by enabling models to acquire new knowledge sequentially while preserving previously learned capabilities. However, achieving this remains challenging, as learning from new data can interfere with prior learning, leading to catastrophic forgetting.
Tokenization is the compression layer between text and every language model, and the field has long assumed that a “good” tokenizer should respect morpheme boundaries and produce subwords that align across languages. This talk examines whether those intuitions hold up for modern LLMs. I present token alignability, a word-alignment-based measure that predicts cross-lingual transfer better than vocabulary overlap, and AlignScore, which assesses morphological plausibility without gold segmentations.
Neural networks are powerful function approximators, but turning that flexibility into scientific discovery often requires a great deal of additional work. In this talk I will discuss two projects applying neural networks to partial differential equations. Although the applications and methods differ, they offer shared lessons about mathematical structure, and optimisation.
This free event is open only to members of the University of Cambridge (and affiliated institutes). Please be aware that we are unable to offer consultations outside clinic hours.
André Freitas (Idiap Research Institute & University of Manchester)
How do language models reason, and what would it take to build AI systems capable of sustained, complex reasoning? This lecture develops the foundations for answering these questions, connecting what we know about the internal workings of language models with methods for evaluating and extending their capabilities.
Mikolaj Kowalski, Jack Franklin, David Kamm - RSEs, ICCS, University of Cambridge
Julia is a just-in-time (JIT) compiled language with a strong focus on scientific computing and performance. It has gained significant popularity in the climate science community, largely due to its flexibility and excellent native support for GPU computing. Over the last year, our team at the Institute for Computing for Climate Science (ICCS) has collaborated closely with the developers of major Julia-based ocean and atmosphere models to optimize their performance on GPU architectures.
Professor Bishop will discuss the rapid advancement of artificial intelligence and its transformative impact across fields including software engineering, weather forecasting, mathematics, scientific discovery, and engineering simulation. The lecture will also highlight an ambitious student rocketry project being undertaken within the Department.
Abstract: The UK’s AI Security Institute (AISI), situated within the Cabinet Office, works to equip government with an evidenced understanding of the risks posed by advanced AI. AISI's Red Team stress-tests frontier safety measures: finding ways misuse safeguards can fail, investigating unintended and harmful model behaviours, and assessing whether protocols intended to keep models under control actually work.
This free event is open only to members of the University of Cambridge (and affiliated institutes). Please be aware that we are unable to offer consultations outside clinic hours.
