Multimodal Neurons in Artificial Neural Networks — LessWrong