Attribution Preservation in Network Compression for Reliable Network Interpretation

Neural Information Processing Systems 

This phenomenon arises due to the fact that conventional network compression methods only preserve the predictions of the network while ignoring the quality of the attributions.