Retraining as Approximate Bayesian Inference: A Decision-Theoretic Framework by Harrison Katz

Summary

A paper by Harrison Katz (arXiv:2603.25480, published March 26, 2026) reframes machine learning model retraining as approximate Bayesian inference under computational constraints. Instead of viewing retraining as routine maintenance on a calendar schedule, the paper treats the deployed frozen model as a belief state that becomes stale as new data arrives. The gap between the continuously updated belief state and the frozen deployed model is termed "learning debt." Retraining decisions are cast as a cost-minimization problem balancing this debt against the computational cost of updating. This yields a decision-theoretic framework for retraining policies, producing evidence-based triggers that replace fixed calendar schedules. The approach also aims to make model governance auditable, since retraining decisions are grounded in measurable evidence rather than arbitrary timing. The result is a principled basis for deciding when to retrain deployed ML models.