The concept of probably approximately correct (PAC) learning has been foundational in computational learning theory.
"The most important thing in science is not so much to obtain new facts as to discover new ways of thinking about them."