Abstract:
This paper extends and improves work of Fortnow and Klivans [2009],
who showed that if a circuit class C has an efficient
learning algorithm in Angluin's model of exact learning via
equivalence and membership queries [Angluin, 1988], then we have the
lower bound EXPNP ⊈ C. We use entirely
different techniques involving betting games [Buhrman et al., 2001] to
remove the NP oracle and improve the lower bound to EXP
⊈ C. This shows that it is even more difficult to
design a learning algorithm for C than the results of Fortnow
and Klivans indicated. We also investigate the connection between
betting games and natural proofs, and as a corollary the existence of
strong pseudorandom generators.
Full Version: