Bayesian Conformal Prediction as a Decision Risk Problem
We present a unified Bayesian decision-theoretic framework for conformal prediction (CP) that integrates Bayesian posterior predictive scores with risk minimization via Bayesian …
🚀Fanyi Kaitlyn Wu
(She/Her)
PhD Student in Decision Making for Complex Systems CDT
Professional Summary
Hello! Welcome! 👋
I am an AI PhD student in Decision Making in Complex Systems CDT led by University of Manchester and University of Cambridge, fully funded by UK Research and Innovation. I am supervised by Dr Michele Caprio and Prof Samuel Kaski.
My current research interest is in AI for Science and uncertainty quantification. Specifically, I am studying Conformal Prediction (CP) and its relationship with other fields (such as Baysian infused CP, CP in medical diagonotics, CP in Astrophyics, etc). I enjoy very much on interdisplinary research and developing theoretical extensions on CP.
My previous backgrounds were in Bsc Physics department at Imperial College London and MPhil in Data Intensive Science at DAMTP, University of Cambridge. My bachelor thesis was about Single Photon Pairs Generation and Detection supervised Dr Stefano Vezzoli in Complex Nanophotonics Group. My master thesis was about Combinatorial Optimisation in Ising Machine Using Optical Parametric Oscillatior System, supervised by Prof Natalia Berloff.
I am also open to industrial experiences and visiting interdisplinary reasearch groups.
Education
PhD Computer Science (AI Focus)
University of Manchester
MPhil Data Intensive Science
University of Cambridge
BSc Physics
Imperial College London
Interests
Uncertainty Quantification
Conformal Prediction
AI for Science
Machine Learning
📚 My Research
We examine Bayesian infused Conformal Prediction (BCP), which merges Bayesian and frequentist methods by using the posterior predictive distribution for the analysis at hand as a non-conformity score in Conformal Prediction (CP).
CP constructs prediction regions with guaranteed marginal coverage properties without distributional assumptions besides exchangeability. These regions contain the true outcome with a pre-specified probability (1-α). Bayesian prediction provides well-calibrated uncertainty when the assumed model is correct but may exhibit poor coverage under model misspecification (M-open perspective).
Our research aims to develop enhanced methodologies that leverage both paradigms’ strengths: incorporating structured prior information from Bayesian statistics while preserving the coverage guarantees of conformal prediction. We also would like to implement variants of CP in different real-world challeges (Physics, Chemistry, Medical diagnotics etc).
Please reach out to collaborate 😃
Featured Publications
Space-time optical diffraction from synthetic motion
Synthetic space-time diffraction from modulating apertures reveals a novel coupling of frequency and momentum, enabling programmable spatio-temporal transformations of light.
Harwood, A. C.
•
•
1 min readRecent Publications
(2025).
THCM-CAL Temporal-Hierarchical Causal Modelling with Conformal Calibration for Clinical Risk Prediction.
EMNLP 2025.
(2025).
Space-time optical diffraction from synthetic motion.
Nat Commun.
Recent News
Trending
✨ Exciting News from EurlPS 2025!
I am glad to attend and present our poster "Bayesian Conformal Prediction as a Decision Risk Problem Using Bayesian Quadrature" at 2025 EIML@[EurIPS](https://eurips.cc/).
Trending
🎉 Presented my first poster at 2025 RSS Conference
Presented poster'Bayesian Infused Conformal Prediction' at 2025 RSS (Royal Statistical Society) in Edinburgh
Trending
📑 2025 ELLIS Cambridge Summer School
Attended and Presented poster'Bayesian Infused Conformal Prediction' at 2025 ELLIS Summer School at Cambridge
Trending
⚛️ First participated Paper pays back
Thrilled to share that the paper I contributed to during my undergraduate in Riccardo Sapienza's lab at Imperial College London has finally been published in Springer Nature …