Learning from What's Right and Learning from What's Wrong

Abstract

The concept of updating (or conditioning or revising) a probability distribution is fundamental in (machine) learning and in predictive coding theory. The two main approaches for doing so are called Pearl's rule and Jeffrey's rule. Here we make, for the first time, mathematically precise what distinguishes them: Pearl's rule increases validity (expected value) and Jeffrey's rule decreases (Kullback-Leibler) divergence. This forms an instance of a more general distinction between learning from what's right and learning from what's wrong. The difference between these two approaches is illustrated in a mock cognitive scenario.