PhD Candidate, University of British Columbia
I am a PhD candidate in Biomedical Engineering at the University of British Columbia, in the de Boer Lab. Almost every model of gene regulation is trained on data that was generated for some other reason: an atlas, a consortium characterisation, whatever happened to get measured. The field has built remarkable models on top of leftovers, and has rarely built the experiment for the model.
I am working on addressing that gap. I design experiments whose purpose is to synthesize the most suitable data for training models, at a throughput worth training on.
Better data is only half of it. Every experiment and dataset carries its own bias, and a model will fit that bias as readily as the biology, so the rest of my work is on ensuring the models learn causal structure rather than the correlations an assay left behind, on a faithful reporting of the model’s performance, and on knowing when to trust a prediction and the interpretation we draw from it. Trust is what makes them worth using.
I am always looking for students to work with. I have supervised six co-op and PhD students in the de Boer Lab, and motivated undergraduates and high-school students are welcome to get in touch.
Vancouver, Canada
Current work
A model is bounded by its data. Our ability to report how much it has learned is bounded by how honestly it was tested. What we can conclude from it is bounded by the tools we read it with. And clinical deployment is bounded by how far we can trust an individual prediction.
Which experiment would teach a model the most? In biology the question is barely asked, because the people generating data and the people modelling it are usually solving different problems. Answering it turns library design into an inference problem rather than a cataloguing exercise.
Architecture, objective, augmentation, and the dozens of small decisions in between. A published model arrives with all of them bundled together, so comparing two models rarely reveals which choice actually carried the result.
A split holds nothing out when related sequences sit on both sides of it, and a model that recalls its neighbours scores like one that understands them. Telling recall apart from reasoning is a prerequisite for every claim anyone makes from a model’s internals.
An aggregate benchmark number says nothing about the case in front of you, and that is the one that matters wherever a model is actually deployed. I work on per-prediction reliability estimates, so that a model can say when it does not know.
Attribution and perturbation methods are instruments in their own right, and largely untested ones. A confident, accurate model read through a broken lens is worse than no model at all.
Topics I am interested in
17 entries
Output
Elsewhere on this site