Convolution as smoothing — LessWrong