I was thinking. What if you rotate an object randomly during training in machine learning. Then the recognition of a rotated object would pose less of a problem.
Then if you rotate some other similar object randomly you will have many predictions to go by. Improving the accuracy then becomes discarding the probable miss classifications. Since they are not in the majority of the cases. Some median of the predictions. Or maybe its better to check the sequence of probability. That is. If you get the same result for the first ?10 predictions.
So from this I guess that the random spin properties of atom objects. Could perhaps in the machine learning sense be sequence sampling probability classifications. That is. If the atom objects are classified to be part of the same atom.