Training

We will need to build a collection of training data and test data. This will make it possible to automatically test the accuracy of new or modified algorithms. For word-by-word methods, the training data and test data will need to include many of the same words.

It will be necessary to be able to store pipeline results. Especially results that receive manual feedback, so that this wouldn't be necessary every time.

It should be possible to save pipelines and sets of pipelines as well.

It may be much more convenient to be able to store internal classifier data (where applicable). Brute-force methods will still need to store the features detected during training.

The pipeline for training saves the features detected to the datastore. The pipeline for classification passes the features to the classifier; the brute-force classifier (DifferencesClassifier) compares them to all of the examples in the datastore (which can be from several images).