It's a new name for computing neural network weights, by analogy to "linear programming" , where the program is optimizing a mathematical loss function with app

It's a new name for computing neural network weights, by analogy to "linear programming" , where the program is optimizing a mathematical loss function with approximate real-number parameters (as opposed to a program that computes a precise result with Boolean /integer logic (such as a web server or a database)). Mathematically, it's analysis on parameters. Confusingly and beautifully, there is also "mechanical differentiation" on the algebra of data of types, another computer application of the same abstract mathematical concept of differentiation: https://en.m.wikibooks.org/wiki/Haskell/Zippers#Mechanical_Differentiation https://en.m.wikibooks.org/wiki/Haskell/Zippers#Mechanical_D... Read about both at http://conal.net/blog/tag/derivative http://conal.net/blog/tag/derivative