We introduce a novel computationally feasible method of exploring protein folding spaces, while working with- out any a priori knowledge of the space, hence over- coming a drawback of previous methods. Since our ap- proach uses random landmark structures without any guarantee of chemical or biological meaning, we use neu- ral networks to extract meaningful features to guide the simulation. The method is described in detail, and its implementation is publicly available. We demonstrate the feasibility of this approach by accelerating the fold- ing of the Trp-Cage miniprotein 20 times.