This page runs the trained model directly in your browser via ONNX Runtime Web -- no server, nothing to keep running. It is a math and machine learning methods study on a large tabular classification problem. No biological, diagnostic, or clinical claim is made anywhere on this page or in this project.
runs entirely client-side -- nothing sent to any serverThe model is a set-equivariant encoder (ISAB / PMA, Lee et al. 2019) pretrained with VICReg self-supervision, finetuned with a Bayes-by-Backprop linear head. This page exports that same model to ONNX and runs it with ONNX Runtime Web using mean posterior weights (a deterministic pass) for a model that works with no backend. The full Bayesian predictive distribution, calibration analysis, and ablation study against simpler baselines are in the full dashboard (run it with Docker, see the footer link) and the source repository.
Class labels are the standard TCGA pan-cancer project categories, used here strictly as category ids for a 32-class classification problem -- not as a diagnostic or clinical statement about any sample.