We show that learning a convex body in $\RR^d$, given random samples from the body, requires $2^{ \Omega(\sqrt{ d/\eps})}$ samples. By learning a convex body we mean finding a set having at most $\eps$ relative symmetric difference with the input body. To prove the lower bound we construct a hard to learn family of convex bodies. Our construction of this family is very simple and based on error correcting codes.

Categories: cs.LG cs.CG

Learning convex bodies is hard