A paper on Neural Network Stability — LessWrong