Research Engineer at Quinze-Vingts Hospital in Paris. Machine learning for ophthalmology, corneal geometry, and the systems that make research reproducible.https://nassimlouissi.com/https://nassimlouissi.com/blog/reading-pretraining-diagnostics/https://nassimlouissi.com/blog/reading-pretraining-diagnostics/What participation ratio, PC1 concentration, and a shared image panel reveal about five training runs—and which questions still need a different experiment.Thu, 01 Oct 2026 00:00:00 GMThttps://nassimlouissi.com/blog/learning-without-labels/https://nassimlouissi.com/blog/learning-without-labels/Follow one image from pixels to representations, calculate a training loss, and see why agreement alone can teach a model nothing useful.Thu, 01 Oct 2026 00:00:00 GMThttps://nassimlouissi.com/blog/five-oct-pretraining-recipes/https://nassimlouissi.com/blog/five-oct-pretraining-recipes/An experimental design walkthrough: full-field MAE, matched patch budgets, diffusion reconstruction, and global–local JEPA, with explicit tradeoffs.Thu, 01 Oct 2026 00:00:00 GMThttps://nassimlouissi.com/blog/from-device-export-to-research-system/https://nassimlouissi.com/blog/from-device-export-to-research-system/A model can accept the right number of inputs and still receive the wrong information. Following one measurement through a system explains why.Thu, 01 Oct 2026 00:00:00 GMThttps://nassimlouissi.com/blog/from-corneal-points-to-zernike/https://nassimlouissi.com/blog/from-corneal-points-to-zernike/A smooth corneal map hides several decisions. Separating measurement, interpolation, and fitting makes those decisions easier to inspect.Thu, 01 Oct 2026 00:00:00 GMThttps://nassimlouissi.com/blog/caching-geometry/https://nassimlouissi.com/blog/caching-geometry/Repeated interpolation becomes a different problem when the sample positions stay fixed. The useful optimization is to identify that invariant precisely.Thu, 01 Oct 2026 00:00:00 GMT