One thing about DL researchers that has always been surprising to me， is that a lot of them have never been exposed to forms of learning other than fitting the parameters of a curve via gradient descent， and are even unable to conceive that there might exist other options