What evidence do you have that this technique is overfitting the training data rather than reading the brain?