Invited Talks

Explainable machine learning is often proposed as a tool to increase the transparency of machine learning systems and to satisfy requirements that might come with the European AI Act. However, to satisfy legal requirements, explanations should not merely appear plausible, but actually be reliable. To investigate when this is the case we introduce a mathematical framework that measures whether explanations are informative. Surprisingly, we find that most of the widely used post-hoc explanations are only informative if the underlying decision function is already simple. This questions the use of post-hoc explanations for complex machine learning models in a legal context. We suggest a path forward by augmenting explanations with concrete, valid interpretations. This can make explanations both formally informative and intuitively useful.

Now that AI has solved a large number of major and long-standing open problems in mathematics, many people have been tempted to say that "mathematics is solved". I shall argue that there is an important sense in which mathematics is far from solved, even if we end up with a black box that is quicker than humans at solving any mathematical problem. I shall also describe some of the work of an automatic theorem proving group I run in Cambridge, and explain what I see as the potential benefits of continuing that work.

Modern Artificial intelligence has progressed under extraordinary abundance: substantial capital, energy and compute, massive model architectures and datasets. These conditions have shaped the systems we have built but have also defined how we measure success. Impressive models that are meant to generalise fail in low resource environments, so what is there to do about it? At Lelapa AI, we develop speech recognition technology for multilingual environments where people routinely mix languages within a single conversation, training data is scarce and compute is constrained. This work exposes two important questions: how well do conventional metrics such as word error rate reflect human judgements of transcription quality? And when can smaller, more efficient models, trained on the right data, outperform larger general-purpose systems on real-world tasks? These questions challenge assumptions about both how we evaluate AI and what it takes to build capable systems. Constraints are not simply obstacles to overcome; they can reveal weaknesses in established methods and open new directions for model development. As AI cannot establish on its own what counts as good. As capabilities and access to compute continue to evolve, the lessons from building under constraint remain valuable: they help us distinguish genuine progress from progress measured by imperfect proxies, and identify what it will truly take to scale.