Control design for multivariable systems is inherently challenging because a control action applied to one channel typically influences multiple transmission paths. Consequently, improving the performance of a selected transmission path may inadvertently degrade the performance of others. To address this fundamental issue, Smith and Wang proposed the disturbance response decoupling (DRD) theorem, which enables the performance of a selected transmission path to remain unchanged while allowing improvements in other paths.
The dominant paradigm in language modeling—scaling next-token prediction with parametric knowledge storage—delivers impressive capabilities but also fundamental limitations: brittle factual memory, inefficient parameters, and myopic reasoning. Progress requires a shift toward external memory and architectures that reason globally before committing to tokens.
Positional encodings are essential for transformer-based language models to understand sequence order, yet their influence extends far beyond simple position tracking. This talk explores the landscape of positional encoding methods in LLMs and reveals surprising insights about how these architectural choices shape model behavior.
We begin with the fundamental challenge: why attention mechanisms require explicit positional information.
Hamiltonian Monte Carlo (HMC) and its variants are among the most widely used algorithms for sampling from probability distributions. Despite their popularity, quantitative convergence guarantees for unadjusted HMC remain limited, especially in divergences that provide strong relative-density control such as KL divergence and Rényi divergence. In this talk, we establish regularization properties for unadjusted HMC via one-shot couplings, which enable Wasserstein convergence guarantees to be upgraded to guarantees in KL and Rényi divergence.
Rohit Bhattacharya, Assistant Professor, Williams College, USA
A recurring question in network studies is whether two connected units resemble each other because one influenced the other (contagion) or because they were alike due to unmeasured background conditions (latent confounding, of which homophily is the canonical case). These are famously hard to separate from a single observed network.
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.
If you would like to participate, please sign up as we will not be able to offer a consultation otherwise. Please sign up through the following link: https://forms.gle/oKKFG78k4CrcE6JK6. Sign-up is possible from June 4 midday (12pm) until June 8 midday or until we reach full capacity, whichever is earlier. If you successfully signed up, we will confirm your appointment by June 10 midday.
Abstract: Explainable AI (XAI) works on providing explanations that justify a model's behaviour or decision. But what is an explanation worth if the user it is meant for cannot understand it? “Social XAI”, a recent interdisciplinary offshoot at the intersection of XAI, dialogue research, and the social sciences (Rohlfing et al. 2021, 2026), shifts the focus to the practice of explaining: the dialogic process through which explanations and their understanding are co-constructed between explainer and explainee.
Professor Yinqing Li, The IDG/McGovern Institute for Brain Research, Tsinghua University & Gurdon Institute, University of Cambridge (Sabbatical Visitor)
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.
If you would like to participate, please sign up as we will not be able to offer a consultation otherwise. Please sign up through the following link: https://forms.gle/HdHM5kKYuxcdRPzr6. Sign-up is possible from June 18 midday (12pm) until June 22 midday or until we reach full capacity, whichever is earlier. If you successfully signed up, we will confirm your appointment by June 24 midday.
