Lifelong Learning Systems Decide Model Reuse or Creation.

For machine learning systems that operate continuously on data streams, effectively managing a collection of specialized "expert" models is critical for lifelong learning. The core challenge is making informed decisions about whether to apply an existing model to new incoming data, determine if a new model is needed, or temporarily postpone a decision. This research proposes a novel decision layer that provides statistical rigor to these choices. It frames the decisions of reusing an expert or spawning a new one as one-sided sequential hypotheses, based on a conditional discrepancy measure. An "indifference zone" is introduced for situations where neither hypothesis has sufficient evidence, leading to a deferral. The framework guarantees finite-time, anytime validity for these decisions. To ensure the system remains responsive to recent data without compromising these guarantees, it employs a restarted e-detector bank. This mechanism efficiently manages error budgets across multiple experts and over time, allowing for robust adaptation in dynamic, unbounded data streams.