On the Probabilistic Learnability of Compact Neural Network Preimage Bounds