The process used in creating the model shall be documented in a model development log ([U]). The development log shall document and justify all key decisions made during the learning process, including the choice of development tool chain (e.g. Tensorflow or pyTorch machine learning platforms), and how those choices impact the performance or robustness of the model.
The model development log may include details of model selection, changes to hyperparameters or changes to trade‐off threshold definitions etc. For example “the threshold level for classification was set to x to ensure that the number of false positives identified in the development data was less than y”. This information could aid developers when trying to develop a model that achieves acceptable false-positive rates.
In order to avoid overfitting and ensure that the model is able to perform in the presence of noise early stopping was employed. When the loss associated with model learning remained with ϵ for 10 iterations training was terminated.
To ensure that the simplest model possible to meet the safety requirements was selected, regularisation is employed, adding a cost to the loss function to penalise the number of neurons in the model.