Exploiting Non-Linear Redundancy for Neural Model Compression