Relationship between training and trained system?

Is there a legal relationship between the training method and a trained model?

Creating a trained machine learning model requires three fundamental parts: a training method, a model and data to train the model with. Once the model is trained and all the parameters of the model are determined, the data and training method are usually no longer needed. The model can be used without knowing how it was trained and with what data. This allows a developer to create and deliver a model in a format where all the parameters can be read without giving away his methods for creating the model. Since the model can exist without the training method, there is no reason that the two must be connected.

The developer will often argue that the training method is part of his expertise and therefore all his own. However, the training method may be important in assessing a model that makes certain mistakes. It is then up to the client and the developer how they want to handle this. The developer can for instance give an explanation without releasing his exact methods. In this way, his expertise remains protected.

In the case of a model that improves itself, the method for making that improvement, the training method, does form part of the model. The user of the model may not receive an explanation from the developer about the exact operation of the training method then, but the developer will have to provide the training method as part of the software that includes the trained model.