We give the first provably efficient algorithm for learning a one hidden
layer convolutional network with respect to a general class of (potentially
overlapping) patches. Additionally, our algorithm requires only mild conditions
on the underlying distribution. We prove that our framewo