They may be unavailable once lacking training samples. In addition, these approaches only use features from the deepest layer and exclude shallow features, which is a kind of information loss. To ...