Parameter Drift

At its core, an AI system consists of a very large number of numbers. These numbers are the adjustment screws that get set while learning from examples. Experts call them parameters or weights. If a system is later retrained again with new data, these numbers shift a bit. If this happens over and over, they gradually move away from their original state, step by step. This very slow, creeping shift is called parameter drift.

When the model answers differently today than it did yesterday

A company thoroughly tests an AI before deploying it. It measures how well it recognizes invoices or sorts customer inquiries. However, this test result only applies to the tested state. If the parameters shift afterward, the result strictly speaking no longer applies. The old test then describes a system that no longer exists in that form.

What’s particularly unpleasant is that parameter drift usually proceeds inconspicuously. The system doesn’t crash and doesn’t report an error. It simply gets a little worse in certain areas while getting better in others. Such deteriorations often only become noticeable once users complain. In credit decisions or medical evaluations, this can become costly or dangerous.

This is why the term is also of interest to regulatory authorities. Anyone operating an AI in a sensitive area must be able to demonstrate that it works reliably. A system whose parameters are constantly moving in an uncontrolled way is hard to document. Many operators therefore deliberately freeze their models and only change them in planned version jumps.

Why the numbers shift in the first place

Training always works according to the same basic principle. The system makes a prediction, compares it with the correct answer, and measures the error. Afterward, all parameters are shifted minimally in the direction that makes the error smaller. A single step barely changes anything. Millions of such steps, however, noticeably change the system.

It becomes critical when the new training data looks different from the old data. The system then aligns itself strongly with the new data and unlearns parts of the old. Experts refer to this as catastrophic forgetting. Parameter drift is the milder, slower variant of the same effect. You can picture it like a compass that shifts by one degree with every touch: after a hundred touches, it points in a noticeably different direction.

There are several countermeasures. You can mix in old training examples so the system doesn’t forget them. You can mathematically restrain important parameters so they are barely allowed to move anymore. And you can train only a small additional layer while the actual model remains unchanged. This last method is particularly widespread in practice today.

The term in product news and model versions

The term is most commonly encountered among companies that operate their own AI systems. They regularly feed the system fresh data from their own operations. In doing so, they continuously monitor whether the hit rate remains stable. If it drops, they switch back to an older, saved version. Such rollbacks are the practical way of dealing with parameter drift.

The topic is also present with well-known chatbots. Users repeatedly report that a model used to solve a certain task better. Sometimes this is due to changed guidelines or safety filters, sometimes it is actually due to retrained parameters. This is precisely why providers give their models version numbers with dates. This allows developers to specifically address a fixed version.

It’s important to distinguish this from a related term. With data drift, the world changes while the model stays the same: customers behave differently, prices rise, new products are added. With parameter drift, conversely, the model itself changes. Both effects lead to declining quality but have different causes. Anyone who confuses them is looking for the error in the wrong place.