> Neural networks find features via gradient descent so humans don't have to encode all of the features. Yes, but what are those features? What did it decide w

> Neural networks find features via gradient descent so humans don't have to encode all of the features. Yes, but what are those features? What did it decide was important? Often, we aren't really sure. We often can't ask the model to justify its decision.