Example Projects

Challenge

Sampling from complex energy landscapes (Boltzmann distributions) is a fundamental challenge in science and ML. Traditional methods like MCMC are often slow to converge, struggle at large scale, and ML-based solutions may lack guarantees of correctness.

Description

This project will explore the use of generative models, specifically normalizing flows and diffusion models, to accelerate high-fidelity sampling from complex Boltzmann distributions. The research will focus on developing and evaluating techniques to overcome the limitations of traditional MCMC methods, aiming for efficient sampling at scale while ensuring high precision. Students will apply these methods to challenging problems in computational chemistry and physics, such as sampling molecular conformations or spin-glass configurations, and benchmark their accuracy in estimating key statistical properties (e.g., moments, free energies).

Skills Required

Probabilistic modelling, MCMC, normalising flows, diffusion models

Skills to be Developed

Generative models, advanced sampling methods

Challenge
Standard foundation models often produce confident but incorrect predictions, limiting their reliability for high-stakes scientific and medical applications.

Description
Investigate and implement state-of-the-art uncertainty quantification techniques (e.g., conformal prediction, Bayesian deep learning, ensembles) for large language models and diffusion models. The project will focus on developing methods to produce well-calibrated confidence intervals and reliable uncertainty estimates for outputs in both BioFM and PatientJourneyFM contexts, enhancing model trustworthiness.

Skills Required
Bayesian methods, conformal prediction, calibration techniques

Skills to be Developed 

Uncertainty estimation, trustworthy AI

Challenge
Despite their enormous popularity, the mechanisms that LLMs use to perform tasks and the related failure modes are poorly understood.

Description
Research methods for incorporating causal knowledge or discovering learnt causal structure within large language models (LLMs) and diffusion models, aiming to improve our understanding of these models, generate mechanistic hypotheses and seamless integration of observational and interventional data during training and inference.

Skills Required
Causality, LLMs, machine learning theory

Skills to be Developed
Causal inference, representation learning

Challenge
Many practical applications rely on fine-tuning predictors on top of pre-trained embeddings, yet a rigorous theoretical framework explaining when this approach is truly optimal compared to end-to-end learning is lacking.

Description
This project aims to develop a theoretical framework, potentially using tools from statistical learning theory and information theory, to understand the conditions under which using pre-trained embeddings is optimal. The students will analyse the interplay between the pre-training data distribution, the downstream task data distribution, the size of the fine-tuning dataset, and model architecture to derive bounds and conditions that guide the choice between a frozen embedding approach (y = f(e(x))) and end-to-end model training (y = f(x)).

Skills Required
Statistical learning theory, information theory, deep learning fundamentals

Skills to be Developed
Theoretical analysis, generalisation theory, embedding models

Challenge
While theoretically appealing, there is an ongoing debate and a lack of clear quantitative understanding of the performance trade-offs associated with enforcing equivariance in neural networks versus allowing them to learn symmetries from data.

Description
This project will develop a mathematical and empirical framework to quantify the sample complexity and performance benefits of equivariant models. The student will investigate how factors like dataset size, the complexity of the symmetry group, and the use of data augmentation affect the performance gap between equivariant and standard architectures. The goal is to provide a clear understanding of when explicit equivariance provides a significant advantage, and when standard models can effectively learn the necessary symmetries from data alone.

Skills Required
Group theory, Statistical learning theory, information theory, deep learning fundamentals

Skills to be Developed
Equivariant neural networks, empirical benchmarking

Description
Design and evaluate a reinforcement learning (RL) or planning-based agent capable of optimizing experimental decisions in a simulated laboratory environment. The goal is to emulate aspects of autonomous scientific discovery by developing an agent that selects actions (e.g., experimental conditions or measurement sequences) to maximize information gain or target outcome efficiency. The project will focus on proof-of-concept systems, such as toy models of chemical or biological processes, where the agent learns optimal strategies for experiment design.

Skills Required
Reinforcement learning fundamentals
Basic probability and decision theory
Familiarity with OpenAI Gym or similar simulation environments

Skills to be Developed
Experimental design and active learning
Policy learning and model-based RL
Integration of physical process knowledge into AI systems
Reproducible simulation pipelines for scientific domains

Challenge
Many domains (genes in DNA, chapters in books) exhibit long-scale hierarchical dependencies that are poorly captured by existing autoregressive foundational models.

Description
Develop and evaluate hierarchical or multiscale masked autoencoder (MAE) architectures for learning rich representations from large biomolecular datasets (e.g., genomics, proteomics), focusing on capturing long-range dependencies and functional motifs.

Skills Required
Deep learning, sequence modelling, MAEs

Skills to be Developed
Multiscale model design, application to biological data

Challenge
While language models have made significant strides in textual reasoning, complex visual reasoning, analogous to question-answering but in the pixel space, remains a frontier. Most models can classify or segment, but struggle with compositional or counterfactual visual tasks.

Description
This project will explore the use of conditional diffusion models as a backbone for visual reasoning. The goal is to move beyond static tasks and develop models that can answer "what if" questions about visual scenes. For example, given an image, the model could be prompted to perform tasks like: "show me this scene if the car were red," "realistically remove this object," or "predict the shadow's position if the light source moved." This involves designing novel visual reasoning benchmarks and developing architectures that can interpret multimodal prompts (e.g., text + masks) to perform complex, compositional image edits that demonstrate a form of visual understanding.

Skills Required
Strong proficiency in Python and PyTorch/JAX, good understanding of deep learning fundamentals, experience with generative models (diffusion models are a plus).

Skills to be Developed
Advanced generative modeling, multimodal fusion architectures, novel benchmark design for AI, conditional image generation, and evaluating abstract reasoning capabilities in vision models.