TOPIC PILLAR / CAUSAL MACHINE LEARNING
Causal Machine Learning & World Models
Predictive accuracy under independent and identically distributed (i.i.d.) observations does not guarantee reliable performance under distribution shifts, policy interventions, or novel counterfactual environments. Our research connects Pearl's Causal Hierarchy (Seeing, Intervening, Imagining) with deep generative modeling and world models: continuous Bayesian causal discovery under unobserved confounders, causal ordering priors, diffusion counterfactuals via semantic abduction, and multi-scale hierarchical causation in complex physical systems.
Articles & Research Explainers
Hierarchical and temporal causation: what changes when causes live at different scales?
Why causal models need to represent both nested levels of abstraction and mechanisms that unfold over time.
A unifying perspective on causal world models
A unifying structural framework mapping visual world models to Pearl's Causal Hierarchy of Seeing, Interpreting, Doing, and Imagining.
Steering optimisation trajectories in diffusion representation learning
Analysing and steering optimisation dynamics reconciles the trade-off between latent disentanglement and pixel reconstruction in diffusion autoencoders.
CO-ALIGN: bias mitigation in diffusion models using Concept Graphs
CO-ALIGN aligns concept graphs between text encoders and visual denoisers to eliminate demographic bias in text-to-image diffusion models.
Emergent tactics in Formula One race strategy via reinforcement learning
RSRL connects real-time telemetry and proprietary simulators to learn emergent pitstop tactics in multi-agent Formula One races.
TRuST: selective fine-tuning for robust concept unlearning in diffusion models
TRuST dynamically identifies concept neurons and applies Hessian-regularised selective fine-tuning to erase harmful concepts from diffusion models.
Causal representation learning with observational grouping for chest X-rays
Observational grouping enforces causal invariance across demographic and scanner environments to eliminate diagnostic shortcuts in chest X-rays.
Diffusion counterfactuals with semantic abduction
Integrating semantic representations into diffusion models through Pearl's causal ladder enables identity-preserving counterfactual image generation.
Causal representation learning via mixture-based priors
Partitioning the latent space into causal factors and exogenous noise enables identifiable causal representation learning without interventional data.
Formula One pitstop strategy via explainable reinforcement learning
Race Strategy Reinforcement Learning (RSRL) optimizes tyre compound choices and pitstop laps in Formula One simulations.
Continuous Bayesian model selection for multivariate causal discovery
A continuous relaxation of Bayesian model selection identifies multivariate directed acyclic graphs without restrictive additive noise assumptions.
Learning hidden variables in a causal order
The paper studies representations whose hidden variables can influence one another.