Terms related to machine learning · 3.3.10
retraining
updating a model that has already been trained by training it again on different data
What this term settles
The entry is short because it is meant to travel. It gives retraining one meaning that can be quoted in a policy without a second interpretation layer.
Who it binds
Nothing here creates an obligation by itself, and the obligation for retraining arrives when the firm writes the term into a control, a specification or a purchase order.
What to do in your practice
Map the places retraining appears in policy, procurement and system records. Where the word is used without the entry behind it, add the citation or replace the sentence.
Evidence an auditor or regulator asks for
- A controlled vocabulary entry for retraining with the citation 3.3.10, an owner and a review date.An auditor can check that retraining is controlled rather than assumed.
- A model record that names the algorithm, the training run and the parameters that changed. The row names retraining.Shows retraining in use in the artefact the firm relies on, not only in a glossary.
- A training log with the data version, the criteria and the result. The record carries the retraining entry and its date.The model record names the vendor product but not the algorithm or the training run, so the term cannot be traced to an artefact. The gap is tested against retraining.
Common gaps this page catches
- retraining is used in the model and training record without a written definition, so two teams can mean different things.
- The entry for retraining is reviewed when someone remembers, not on a cadence, so it drifts from the artefact it describes.
What to prepare before the amendment lands
Add the entry to the register, give it an owner and a review date, and test it against one live artefact this week. The test is the evidence; the register row alone is not.
How to test it here
Take one live artefact rather than a template, find the sentence or field where retraining is used, and put this page beside it. Record the answer, the artefact and the date, because that record is the evidence and a glossary entry on its own is only an intention.
What this page does not claim
The page does not reproduce the standard, certify the firm or predict the wording the amendment will use; it states retraining at the claim level "definition" and stops where the public record stops.
Common questions
- What does retraining mean in ISO 22989?
- updating a model that has already been trained by training it again on different data
- What evidence should we keep for 3.3.10?
- A controlled vocabulary entry for retraining with the citation 3.3.10, an owner and a review date. A model record that names the algorithm, the training run and the parameters that changed. The row names retraining.
- Is retraining a requirement?
- No: retraining is a definition or clause of a terminology standard, and it becomes binding only when a firm writes it into a policy, a contract or a procedure, or when another standard incorporates it.
Related units
- 3.3.9 reinforcement learning
- 3.3.11 semi-supervised machine learning
- 3.3.12 supervised machine learning
Source. ISO/IEC 22989:2022, Clause 3.3.10, Terms and definitions. Open the ISO public entry. Retrieved 21 October 2026.
https://www.iso.org/obp/ui/en/#iso:std:iso-iec:22989:ed-1:v1:en
ISO catalogue record 88145, the amendment status, stage and dates. Catalogue record 74296, the base standard.
Claim level: definition. The definition is in the public preview. This page states it in our own words and does not reproduce the text.
Last updated: 2026-10-21 · Markdown version · Check your vocabulary