M-Rule: An Enhanced Deep Taylor Decomposition for Multi-model Interpretability

Recent intelligent decision systems based on machine learning have achieved superhuman performance by learning rich knowledge from massive amounts of data through complex deep network models. However, interpretability remains a significant challenge in artificial intelligence, as state-of-the-art decision-making frameworks still struggle with unexplained decisions. This paper conducts interpretability research from the dimension of feature contribution. The core idea is to provide a visual explanation that shows which important areas of the input guide the model to make the final decision and helps users understand the reasons for the model decision. For better redistribution and explanation of the correlation between the input and the final output, we first improve the Deep Taylor Decomposition(DTD) process and proposed an enhanced relevance propagation method called “M-Rule”. This method is suitable for use in multi-level decision-making models, including DNNs, CNNs, LSTMs, and more. Therefore, it is considered to be a multi-model interpretability approach. Then We evaluate the interpretability improvement using M-Rule on the ImageNet dataset, and address the small sample problem by replacing the region with the largest contribution in each image to achieve data enhancement and identify the regions of interest (ROIs) in COVID-19 lung CT images. The results demonstrate that the model trained in this paper using interpretation-based augmented data has a significant performance improvement compared to other classification models. With the continuous advancement of machine learning technology, we believe that interpretability will eventually produce an autonomous system capable of making human-comprehensible decisions in the future.