Machine learning models are defined by both the architecture (which may be generic) and the training data. Specifically, the same model may be trained to achieve desired results from the same training data but with different outcome labels.
A lot of traditional algorithm engineering and consequently patents have covered specific methods of analysis. As an example, assume that a patent claim says "Process the data and output a classification by comparing the mean value to a defined threshold."
A neural network (or other "machine learned" algorithms) may perform this step through a generic approximation of the decision boundary. Specifically they will at no point define the threshold in the model architecture but the model parameters will be optimized based on data that may have been labeled by comparing a mean value to a defined threshold. This will not be evident by looking at the implementation of the model, since the training dataset is not part of the final product. But it may be suggested by testing the model performance on such a dataset and seeing how it approximates the patented method.
To my understanding a desired outcome is not itself patentable, while the method of achiving the outcome can be.
Generally, will a generic approximation of a patented method constitute infringement of the method?