Transfer learning adapts a model trained on one task to a related task, which is how most organisations reuse a foundation model.
What the base framework establishes
The clause covers the reuse of a model that was already trained. The base model carries patterns from its original data, and the organisation adjusts it for its own task, either by fine-tuning or by using it as a fixed feature extractor. The framework already uses the idea in the system overview, where a pre-trained model is one trained when obtained. The governance consequence is that the record must describe two datasets and two purposes: the original training and the adaptation. The clause text is not in the public sample, so this page describes the concept and its evidence rather than the clause's words.
Why it matters to an AI management system
Transfer learning hides a supply chain. The organisation may know its own fine-tuning data and know nothing about the base model's data, licence or evaluation. A description that names only the fine-tune is incomplete, and the amendment's generative material makes that gap more common because foundation models are usually consumed this way.
The public draft summary names no change to 7.7 Transfer learning. The final text could still touch it.
Evidence a reviewer samples
| Category | Artefact |
|---|---|
| Base model record | The provider, version, licence and any published evaluation. |
| Adaptation record | The fine-tuning data, the method and the purpose of the adaptation. |
| Two-purpose note | The original task and the new task, with the difference stated. |
One question to prepare for. What do you know about the data the base model was trained on?
Common gaps
- The base model version is unpinned and changes under the application.
- The provider's training data and licence are not recorded.
- The fine-tuning data is small and unrepresentative and nobody says so.
Prepare before publication
Pin the base model version, record its licence and published evaluation, and describe the adaptation separately from the base.
Dates that matter
No separate date attaches to 7.7; the record dates are the stage date, the planning estimate and the CEN expectation stated above.
Cross-references
6.1 Overview of the machine learning system · 6.3 Model · 8.4 Modelling
Source
Source for 7.7: ISO/IEC 23053:2022, clause 7.7. The public sample stops before 7.7, so this page relies on the public contents and the public summary and does not quote the clause. The two official records are linked on this page so any line can be checked.
Not legal advice; does not certify. A reading of the framework and the evidence a reviewer commonly samples, not a substitute for the standard or your auditor's rules.