I have a few funny analogies that I think kind of work. 1. "gradient descent" is like tuning a guitar by ear and listening to the beat frequencies ("loss") and

I have a few funny analogies that I think kind of work. 1. "gradient descent" is like tuning a guitar by ear and listening to the beat frequencies ("loss") and then decreasing these by tuning a string up or down. 2. the best I can come up with for "backpropagation" is to imagine a clever device that can tirelessly optimize a Rube Goldberg machine for you but as a science, not an art.