Depending on how you define “learning” (there are different basic frameworks like PAC learning) there’s a number of analyses that show learning and compression

Depending on how you define “learning” (there are different basic frameworks like PAC learning) there’s a number of analyses that show learning and compression are equivalent. Even quite broadly, Bayesian methods can be interpreted as a rate-distortion problem from Information Theory, which is an approach to lossy compression