A New Modal Autoencoder for Functionally Independent Feature Extraction