Preregistration of Machine Learning Research

Abstract

It is interesting to note that human intelligence thrives on what Peirce called abductive inferences (Peirce and Turrisi 1997, 241-56), which are neither inductive nor deductive. Abductive inferencing basically entails an informed guess as to the explanation of a set of observations. Building on Peirce, scientific research can be framed as starting with an abduction based on observation, generating an explanation (theory) from which a hypothesis (prediction) is deduced about subsequent observations, after which the prediction can be inductively tested against new observations. Building on Popper’s theory of falsification,1 hypotheses should be developed in a way that enables the rejection of the explanation – not merely its verification. A theory that explains why all swans are white should not just be verified by detecting ever more white swans, but tested against its potential falsification by searching for black swans.