A Question That Kept Getting Bigger.
It started with a simple, almost stubborn idea: can a machine know how confident it should be? That question about uncertainty never left — it just found new places to live: in word embeddings, in hypergraphs running inside eBay's production stack, in retinal scans, and now in whether a language model understands the culture of the person talking to it.
How One Question Evolved.
Not a career timeline — that's a different story. This is the arc of the ideas themselves: the same instinct about uncertainty, followed through five very different disguises.
An unsupervised idea about ranking documents by difficulty — inspired, oddly, by Super Mario level design — becomes a PhD in latent probabilistic topic discovery, and a Best Paper award at the Beijing–Hong Kong Doctoral Forum.
Word embeddings stop being black boxes. Conceptual spaces, von Mises–Fisher priors, embeddings framed as principled Bayesian estimation — published at ACL, AAAI and IJCAI.
Multi-modal hypergraphs ship inside eBay's recommendation stack; a transformer model goes live inside BT's customer systems. The lab work starts paying rent in the real world.
Bayesian uncertainty modelling, with Thomson Reuters Labs, asks whether AI should be willing to say "I don't know." The same instinct turns retinal scans into a biomarker for whole-body health — featured on Forbes.
CALM teaches language models cultural self-awareness. Policy work with the Royal Society, and a Ministry of Defence project with the Alan Turing Institute, ask what responsible AI actually requires.
The Same Idea, Four Disguises.
Read top to bottom and you're reading the arc above, slowed down.
Probabilistic Foundations
Every model I've built since 2009 rests on the same conviction: intelligence isn't about being right, it's about knowing how unsure you are. My PhD work on latent topic discovery grew into a decade of Bayesian non-parametric models and approximate inference — the mathematical bedrock that lets AI reason honestly about what it doesn't know, in domains where a falsely confident answer is worse than none at all.
Word Embedding as Maximum A Posteriori Estimation
Word and Document Embedding with vMF-Mixture Priors on Context Word Vectors
Nonparametric Topic Modeling using Chinese Restaurant Franchise with Buddy Customers
BAKER: Bayesian Kernel Uncertainty in Domain-Specific Document Modelling
Multi-modal Learning at Scale
Real data doesn't arrive as clean text — it's images, clickstreams and structured metadata, all at once. I build hypergraph models that capture relationships across these modalities simultaneously, work that moved from a whiteboard to production traffic at eBay and British Telecom, and picked up an eBay Top-3 Leaders' Choice Award along the way, selected out of 2,000+ global submissions.
Click-Through Rate Prediction with Multi-Modal Hypergraphs
Unified Transformer Multi-task Learning for Intent Classification with Entity Recognition
Inferring Prototypes for Multi-Label Few-Shot Image Classification with Word Vector Guided Attention
GAMED: Knowledge Adaptive Multi-Experts Decoupling for Multimodal Fake News Detection
Digital Health & Oculomics
The eye turns out to be a remarkably honest witness to what's happening in the rest of the body. My work in Oculomics fuses computer vision with medical imaging to surface early biomarkers for cardiovascular and systemic disease — research that Forbes described as turning eye exams into a gateway to whole-body wellness, and that leans on Pillar I's uncertainty modelling to make medical predictions AI is willing to stand behind.
The Eye as a Window to Systemic Health: A Survey of Retinal Imaging from Classical Techniques to Oculomics
Explainable Depression Detection with Multi-Modalities Using a Hybrid Deep Learning Model on Social Media
Would You Trust an AI Doctor? Building Reliable Medical Predictions with Kernel Dropout Uncertainty
Retinal–Lipidomics Associations as Candidate Biomarkers for Cardiovascular Health
Responsible AI & Governance
As LLMs get folded into everyday decisions, "it works on the benchmark" stops being good enough. My recent work makes models culturally self-aware and interrogates the bias baked into their architecture and training data — the same questions that led the Royal Society to pair me with a UK Home Office policymaker, and that now underpin a Ministry of Defence project on sensemaking under uncertainty with the Alan Turing Institute.
CALM: Culturally Self-Aware Language Models
Bias in Language Models: Interplay of Architecture and Data?
Evaluating Large Language Models on Health-Related Claims Across Arabic Dialects
Uncertainty Modelling in Under-Represented Languages with Bayesian Deep Gaussian Processes
One Foundation, Three Frontiers.
Hover a node. Nothing here is a coincidence — the uncertainty methods from Pillar I are load-bearing for everything downstream.
Hover or tap a node to see how it connects to the rest of the work.
The Work Doesn't Stop at Publication.
A handful of real papers that cite this work — sourced from Semantic Scholar — chosen to show how far back some of it reaches.
A Comparative Study of Neural Sinkhorn Topic Models Based on Different Word Embeddings (2025)
Builds on Jointly Learning Word Embeddings and Latent Topics (SIGIR 2017), comparing it against newer optimal-transport-based topic models.
8-year gapMulti-graph Collaborative and Global Hypergraph Sampling for Multimodal Recommendation (2026)
Directly extends the eBay-deployed Click-Through Rate Prediction with Multi-Modal Hypergraphs (CIKM 2021) with a new sampling strategy.
5-year gapDep-LLM: Training-Free Depression Diagnosis via Evidence-Guided Structured Multi-factor Reasoning (2026)
Cites the most-cited paper in this whole archive — Explainable Depression Detection with Multi-Modalities (2021, 175+ citations) — while exploring LLM-based diagnostic reasoning.
5-year gapDeepQFM: A Deep Learning Based Query Facets Mining Method (2023)
Builds on Web Query Reformulation via Joint Modeling of Latent Topic Dependency (TOIS 2015) to mine query facets with deep learning.
8-year gap