A Formal Proof of PAC Learnability for Decision Stumps
share
We present a machine-checked, formal proof of PAC learnability of the concept class of decision stumps. A formal proof has every step checked and justified using fundamental axioms of mathematics. We construct and check our proof using the Lean theorem prover. Though such a proof appears simple, a few analytic and measure-theoretic subtleties arise when carrying it out fully formally. We explain how we can cleanly separate out the parts that deal with these subtleties by using Lean features and a category theoretic construction called the Giry monad.
READ FULL TEXT
