Verifying Safety of Neural Networks from Topological Perspectives