Effect of Multimodal Metadata Augmentation on Classification Performance in Deep Learning

The impact of Multimodal Metadata Augmentation (MMA) on data classification accuracy, is explored focusing on datasets such as PathMNIST, RetinaMNIST, and CIFAR10. Experimental results reveal that MMA can significantly improve classification accuracy, with the extent of improvement dependent on dataset characteristics, such as data complexity, sample size, and variability. The study considered various scenarios, ranging from the worst-case accuracy improvement to the best-case scenario, depending on how classification errors are distributed among classes. Theoretical estimations align closely with experimental outcomes, demonstrating the potential of MMA for practical applications. These findings enhance our understanding of MMA’s potential in Explainable Artificial Intelligence and offer valuable insights into dataset quality assessment.