In image processing, I was motivated by Keras Image Generator. Use of Image Generator pumped my Cats ad Dogs classifier from 50% accuracy to 87%. Would it not be useful to identify what set of transformations (like rotation, slicing, cropping etc) makes the test data more predictable? We can create a model( transformation suggestor model) and teach it to suggest the type of transformation(s) that is best suited to increase the accuracy of prediction of the original classifier(classifier model) on any test data . We could put the new data that needs to be classified into the transformation suggestor model first, perform the recommended suggestions and the try out classifier model with the new data. If the classifier classifies it accurately add the new data to the test repository or toss it , if found to be non- classifiable both manually and by the model.