I am trying to figure this world out ^-^
I'm a Machine Learning Engineer at NextGen Federal Systems, where I work on agentic AI and ML systems day-to-day. I am interested in building ML systems on physical signals, where the ground truth is not perfect, the failure modes matter more than the averages, and the system has to be defensible to a domain expert. That is what I did for my undergraduate research in multimodal medical imaging at West Virginia University and what I do now in deep-learning weather models, and what I want to keep working on in graduate school.
Data-driven weather models outperform operational NWP forecasts on aggregate skill scores but remain inexplicable to the forecasters who have to act on them, especially on localized extreme events. Does the inexplicability come from L2 training, GNNs' spectral bias, or training on smoothed reanalysis data?
Of the three, I think the loss function is where I'd want to start. It's something you can iterate on without retraining a foundation model, and it's where over-smoothing and the explainability gap intersect.