about

Walter Nelson AI-generated headshot

I'm an ELLIS PhD student at IST Austria supervised by Francesco Locatello, co-supervised by Arthur Gretton at the Gatsby Computational Neuroscience Unit, University College London.

My recent work focuses on the identifiability of representations in modern machine learning models, particularly foundation-scale models. I'm motivated by problems in biology that offer the chance to gain insights into the mechanisms of human disease and behaviour, and in medicine with the potential for real-world impact on patient care.

Previously, I was a full-time staff data scientist at the Centre for Data Science and Digital Health (CREATE) at Hamilton Health Sciences in Hamilton, Canada. During my undergraduate studies, I worked in the lab of Anna Goldenberg at The Hospital for Sick Children in Toronto. In the past, I've also consulted on software engineering, machine learning engineering, and data science projects ranging from back office finance to consumer-facing mobile and web applications.

employment

  • Research Consultant
    Chan Zuckerberg Initiative Foundation
    March 2025 - June 2025
  • Junior Data Scientist
    Centre for Data Science & Digital Health, Hamilton Health Sciences
    2019 - 2024

education

  • Doctor of Philosophy, Computer Science
    Institute of Science and Technology Austria
    2024 - present
  • Master of Science, Statistics
    University of Toronto
    2021 - 2024
  • Honours Bachelor of Science, Bioinformatics and Computational Biology & Neuroscience
    University of Toronto
    2015 - 2019

selected publications

  • Toward Identifiable Sparse Autoencoders
    ICML, 2026 [arxiv]
    Walter Nelson, Theofanis Karaletsos, Francesco Locatello
  • Statistical and structural identifiability in representation learning
    ICLR, 2026 [arxiv]
    Walter Nelson, Marco Fumero, Theofanis Karaletsos, Francesco Locatello
  • Detecting irregularities in randomized controlled trials using machine learning
    Clinical Trials, 2024 [doi]
    Walter Nelson*, Jeremy Petch*, Jonathan Ranisau, Robin Zhao, Kumar Balasubramanian, Shrikant Bangdiwala
  • Optimizing warfarin dosing for patients with atrial fibrillation using machine learning
    Scientific Reports, 2024 [doi]
    Jeremy Petch, Walter Nelson, Mary Wu, Marzyeh Ghassemi, Alexander Benz, Mehdi Fatemi, Shuang Di, Anthony Carnicelli, Christopher Granger, Robert Giugliano, Hwanhee Hong, Manesh Patel, Lars Wallentin, John Eikelboom, Stuart Connolly

open source & side projects

  • SanteMPI Record Linkage Configuration Optimizer
    Through our work at CREATE, I engineered a plugin for optimizing the parameters of the SanteMPI patient record linkage algorithm using Bayesian optimization. Sante software underpins national deployments in Tanzania and Myanmar.
  • Compare Concordance
    I authored a Python port of the compareC R library for doing statistical inference about correlated right-censored c-indices, which commonly arise in the comparison of predictive survival models. The package is available on PyPi.

research

Now. Lately I've been interrogating the question of when a machine learning model is stable—or, in the language of statistics and causality, when its internal representations are identifiable. In [1], we make rigorous two notions of identifiability that recent literature has tended to conflate: statistical identifiability, the fact that a representation is uniquely learned from the data by a given model; and structural identifiability, the fact that the representation aligns with some ground-truth component of the data-generating process. It's a fairly theoretical paper, but we empirically validate some of its implications for real self-supervised models. In particular, the theory motivates using linear independent components analysis to recover a disentangled basis for the representations of self-supervised models; doing so in a foundation model for biological images verifiably disentangles batch effects from biological signal, improving out-of-distribution generalization. In [2], we investigated the statistical identifiability of sparse autoencoders, a common interpretability tool for LLMs. Drawing parallels to the classical dictionary learning and sparse coding literature, we find that SAEs underperform on identifiability, and trace this to a “dark triad” of SAE characteristics that fortunately admit simple architectural fixes. On synthetic data generated according to the linear representation hypothesis, these fixes uniformly improve performance and identifiability; on real-world data the results are more mixed, but still state-of-the-art on a number of metrics. I'm particularly excited about extending identifiability theory to interrogate, explain, and falsify parts of the so-called “Platonic representation hypothesis.” True to my roots, I'm still working on biology too, with a focus on new methods for underdetermined inverse problems.

Biology. I spent a lot of time thinking about how to apply statistical learning to bioinformatics tasks. For example, in [3], we evaluated a number of methods and asked: does it matter what families of algorithms you use for machine learning on graphs in biology? The short answer: probably not too much. More recently, I've worked on Mendelian randomization, a technique for doing causal inference where genotype serves as an instrumental variable. In [4], we looked hard at the data-generating process for SNP data, and developed a clever way to use the raw probe intensities to get more robust estimates of mitochondrial copy number, providing evidence for its causal role in dementia. In [5], we used machine learning to curate and extract features from retinal fundus images for downstream Mendelian randomization analyses, investigating the link between peripheral vasculature complexity and cardiometabolic diseases. I've also looked at how the standard toolkit for machine learning, like GPU acceleration, can be applied to traditional statistical techniques for large-scale datasets [6].

Health. In my previous positions, most of my work was looking at machine learning for health. A central question we asked in several reviews was: how should we evaluate a machine learning model in the clinical setting, and when can it be trusted [7], [8], [9]? We were early advocates for robust, prospective evaluation of machine learning models as the only credible way to establish their utility in a health care setting. I've collaborated closely with clinicians to build models such as one that predicts whether a patient's liver cancer will recur after transplant [10], and one that predicts whether their peripheral vasculature is showing signs of poorly-managed or undiagnosed cardiometabolic disease [11]. I've developed new methods for linking records in patient databases, a key problem in real-world electronic medical record deployments [12], and contributed to software for deidentifying them to support health care research [13]. We developed a theoretically-justified set of algorithms for applying reinforcement learning to healthcare, and applied them to a difficult problem in medication management: namely, the dosing of warfarin [14]. In [15], we again applied causal principles to our knowledge of the data-generating process to develop a grounded evaluation scheme for our model retrospectively. I've also worked on machine learning for health research administration tasks: we evaluated cutting-edge anomaly detection methods for clinical trials, showing that complicated approaches that appear promising in machine learning benchmarks do not significantly outperform simple baselines in real data [16], [17], [18].

References
  1. W. Nelson, M. Fumero, T. Karaletsos, F. Locatello. "Statistical and structural identifiability in representation learning." ICLR, 2026.
  2. W. Nelson, T. Karaletsos, F. Locatello. "Toward Identifiable Sparse Autoencoders." ICML, 2026.
  3. W. Nelson, M. Zitnik, B. Wang, J. Leskovec, A. Goldenberg, R. Sharan. "To embed or not: network embedding as a paradigm in computational biology." Frontiers in Genetics, 10, 381, 2019.
  4. M. Chong, P. Mohammadi-Shemirani, N. Perrot, W. Nelson, R. Morton, et al. "GWAS and ExWAS of blood Mitochondrial DNA copy number identifies 71 loci and highlights a potential causal role in dementia." eLife, 11, e70382, 2022.
  5. A. Villaplana-Velasco, N. Perrot, Y. Hang, M. Chong, E. Trucco, et al. "Mendelian randomization study implicates inflammaging biomarkers in retinal vasculature, cardiovascular diseases, and longevity." Science Advances, 11(43), eadu1985, 2025.
  6. M. Di Scipio, M. Khan, S. Mao, M. Chong, C. Judge, N. Pathan, N. Perrot, et al. "A versatile, fast and unbiased method for estimation of gene-by-environment interaction effects on biobank-scale datasets." Nature Communications, 14(1), 5196, 2023.
  7. J. Petch, J. P. T. Bortesi, W. Nelson, S. Di, M. H. Mamdani. "Should I trust this model? Explainability and the black box of artificial intelligence in medicine." Artificial Intelligence for Medicine, 265–273, 2024.
  8. J. Petch, S. Di, W. Nelson. "Opening the black box: the promise and limitations of explainable machine learning in cardiology." Canadian Journal of Cardiology, 38(2), 204–213, 2022.
  9. S. Muralitharan, W. Nelson, S. Di, M. McGillion, P. J. Devereaux, N. G. Barr, et al. "Machine Learning–Based Early Warning Systems for Clinical Deterioration: Systematic Scoping Review." Journal of Medical Internet Research, 23(2), e25187, 2021.
  10. T. Ivanics, W. Nelson, M. S. Patel, M. P. A. W. Claasen, L. Lau, A. Gorgen, P. Abreu, et al. "The Toronto Postliver Transplantation Hepatocellular Carcinoma Recurrence Calculator: A Machine Learning Approach." Liver Transplantation, 28(4), 593–602, 2022.
  11. R. Shah, J. Petch, W. Nelson, K. Roth, M. D. Noseworthy, M. Ghassemi, et al. "Nailfold capillaroscopy and deep learning in diabetes." Journal of Diabetes, 15(2), 145–151, 2023.
  12. W. Nelson, N. Khanna, M. Ibrahim, J. Fyfe, M. Geiger, K. Edwards, J. Petch. "Optimizing Patient Record Linkage in a Master Patient Index Using Machine Learning: Algorithm Development and Validation." JMIR Formative Research, 7, e44331, 2023.
  13. C. Moore, J. Ranisau, W. Nelson, J. Petch, A. Johnson. "Pyclipse, a library for deidentification of free-text clinical notes." arXiv:2311.02748, 2023.
  14. J. Petch, W. Nelson, M. Wu, M. Ghassemi, A. Benz, M. Fatemi, S. Di, et al. "Optimizing warfarin dosing for patients with atrial fibrillation using machine learning." Scientific Reports, 14(1), 4516, 2024.
  15. M. Fatemi, M. Wu, J. Petch, W. Nelson, S. J. Connolly, A. Benz, A. Carnicelli, et al. "Semi-Markov Offline Reinforcement Learning for Healthcare." Conference on Health, Inference, and Learning (CHIL), 119–137, 2022.
  16. W. Nelson, J. Petch, J. Ranisau, R. Zhao, K. Balasubramanian, et al. "Detecting irregularities in randomized controlled trials using machine learning." Clinical Trials, 17407745241297947, 2024.
  17. W. Nelson, J. Ranisau, J. Petch. "Do Ensembling and Meta-Learning Improve Outlier Detection in Randomized Controlled Trials?" Machine Learning for Health (ML4H), 2023.
  18. J. Petch, W. Nelson, S. Di, K. Balasubramanian, S. Yusuf, P. J. Devereaux, et al. "Machine learning for detecting centre-level irregularities in randomized controlled trials: A pilot study." Contemporary Clinical Trials, 122, 106963, 2022.

contact

I can be reached by email at .