Is there any Need to Mention Induction?


Induction is the process by which seen data becomes the basis for predicting unseen data. There has long been a desire to explain the procedure in a context-free way. But Hume's circularity problem and the no-free-lunch theorems seem to suggest the impossibility of any context-free mechanism. Machine Learning takes the position no such mechanism exists. But an alternative comes from Epistemology. Popper's falsificationist theory holds there is a general mechanism, but that it does not perform induction. Inductive effects arise implicitly, through pursuit of a non-inductive goal. Less plausibly, the mechanism is taken to be uninformed exploration of hypotheses. But as the paper shows, Popper's solution can be reworked using information theory. Increasing the informational efficiency with which representations predict seen data can be shown to produce inductive effects. With representation optimization taking the place of hypothesis-search in the argument, it becomes possible to explain induction in a context-free way.

Back to Table of Contents