LioNets: Local Interpretation of Neural Networks through Penultimate Layer Decoding