Finding Interpretable Class-Specific Patterns through Efficient Neural Search