Visualizing Neural networks, how to blame the bias — LessWrong