Two tool terms are defined: backpropagation, the method that adjusts a neural network from its output error, and the learning rate, the step size that controls how fast training moves.
What the base framework establishes
The clause keeps the tool vocabulary small. Backpropagation is described as the way the error at the output layer is used to adjust the weights of the connections in the layers before it. The learning rate is the step size of a gradient method, and the note explains that it decides whether and how quickly a model converges, which is why it is treated as an important setting for neural networks. The point is not to teach optimisation. It is to give a shared name to the two settings that most often explain why two runs of the same experiment produced different models.
Why it matters to an AI management system
When a model is retrained and behaves differently, the governance record should be able to say whether the data, the architecture or a training setting changed. These two terms make that conversation possible without a data scientist in the room.
The public draft summary names no change to 3.2 Tools. The final text could still touch it.
Evidence a reviewer samples
| Category | Artefact |
|---|---|
| Training configuration | The recorded hyperparameters, including the learning rate and its schedule. |
| Run comparison | The difference between the prior and current training run, with the changed setting named. |
| Toolchain record | The framework and library versions used for the run. |
One question to prepare for. If the model changed behaviour last month, could you say which setting changed?
Common gaps
- The training configuration lives only in a notebook and cannot be reproduced.
- A retrain changes several settings at once, so the cause of a behaviour change cannot be isolated.
- The model vendor's toolchain version is not recorded in the system description.
Prepare before publication
Freeze the training configuration into the model record and require a run comparison note for any retrain that changes a setting.
Dates that matter
No separate date attaches to 3.2; the record dates are the stage date, the planning estimate and the CEN expectation stated above.
Cross-references
6.5 Tools · 8.4 Modelling · 8.7 Operation
Source
Source for 3.2: ISO/IEC 23053:2022, clause 3.2. The statement is read from the public base sample at 3.2 and is written in our words. 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.