Shallow2Deep: Restraining Neural Networks Opacity Through Neural Architecture Search

Recently, the Deep Learning (DL) research community has focused on developing efficient and highly performing Neural Networks (NN). Meanwhile, eXplainable AI (XAI) making Machine (ML) methods interpretable transparent, seeking explainability. This work is a preliminary study applicability of Architecture Search (NAS) (a sub-field DL looking for automatic design NN structures) in XAI. We propose Shallow2Deep, an evolutionary NAS algorithm that exploits local variability to restrain opacity DL-systems through architectures simplification. Shallow2Deep effectively reduces complexity – therefore their while reaching state-of-the-art performances. Unlike its competitors, promotes localised structures NN, helping reduce opacity. The proposed analyses role design, presenting experimental results show how this feature actually desirable.