Correction Cascades
A correction cascade occurs when a model is built to learn small corrections on top of another model’s output, rather than solving the problem directly.
The pattern emerges when model A exists for problem A, but a slightly different problem A′ is needed. It’s tempting to learn a quick correction model that takes A’s output and adjusts it. But this creates a dependency: any improvement to model A now requires re-evaluating the correction model.
The costs compound when corrections are stacked, a model for A′′ built on top of the A′ correction, and so on. This creates an improvement deadlock: improving any individual component may degrade system-level performance.
Mitigation: either augment the original model to handle the variations directly, or accept the cost of building a separate model for each problem variant.
Related: 05-atom—cace-principle, 05-molecule—ml-technical-debt-taxonomy