Robust and Explainable AI: Auto-Augment with Label Preservation and Saliency Parameters

Artificial intelligence (AI) relies on models that are reliable and simple to comprehend. Generative AI is a subset of AI as it can generate something new and unique from random noise or existing data inputs regarding an image, text, and data. In auto-augmentation, a generative AI can help developers better plan and estimate work. This research covers two critical areas of AI research: auto-augmentation and explainability. Additionally, our approach makes machine learning models more accessible to read by maintaining the purity of labels and adding saliency factors for better performance. The data is changed using auto-augmentation methods to improve model adaptation during training. These changes could alter ground truth labels, impacting how well the model works. We offer a label preservation way to rectify issues, ensuring data enhancement processes maintain label consistency. Applying the suggested method's saliency parameters makes it easier to understand how the model's forecasts work, which increases their dependability and openness. Using intermediate layer models without knowing anything about the domain makes hard-positive cases that keep the original labels. It improves performance across disciplines. Seeking model explainability is done with parameter saliency maps. It makes it easier to understand how models behave by finding and studying the network factors that lead to bad decisions. Using a ResNet18 classifier, the suggested method is tested thoroughly on the CIFAR100 dataset.