# Explainable Artificial Intelligence: Understanding, Visualizing and Interpreting Deep Learning Models

**Authors:** Wojciech Samek, Thomas Wiegand, Klaus-Robert Müller  
**arXiv:** [1708.08296](https://arxiv.org/abs/1708.08296)  
**Date:** 2017-08-28  
**Summary page:** https://www.summarizepaper.com/en/paper/1708.08296v1/

## Key points

- Recent developments in interpretability in artificial intelligence are discussed
- AI systems can perform complex tasks at or beyond human-level performance
- Lack of transparency and interpretability is a major drawback of these models
- The authors emphasize the need for methods to visualize, explain, and interpret deep learning models
- Two approaches for explaining predictions of deep learning models are presented: sensitivity analysis and decomposition of decisions into meaningful terms
- These approaches improve interpretability by providing insights into decision-making processes
- Three classification tasks were performed to evaluate these approaches
- The paper highlights the importance of interpretability in artificial intelligence

## Summary

The paper titled "Explainable Artificial Intelligence: Understanding, Visualizing and Interpreting Deep Learning Models" by Wojciech Samek, Thomas Wiegand, and Klaus-Robert Müller discusses the recent developments in the field of interpretability in artificial intelligence. With the advancements in deep learning methodology and the availability of large databases, AI systems are now able to perform complex tasks at or even beyond human-level performance. This progress can be observed in various domains such as image classification, sentiment analysis, speech understanding, and strategic game playing. However, a major drawback of these highly successful machine learning and artificial intelligence models is their black box nature. These models lack transparency as they do not provide any information about how they arrive at their predictions. This lack of interpretability can be problematic, especially in medical applications where it is crucial to understand the reasoning behind AI system's predictions. To address this issue, the authors emphasize the need for methods that can visualize, explain, and interpret deep learning models. They summarize recent developments in this field and advocate for more interpretability in artificial intelligence. The paper also presents two approaches for explaining predictions of deep learning models. The first approach computes the sensitivity of the prediction with respect to changes in the input. This method helps identify which parts of the input have a significant impact on the model's decision-making process. The second approach decomposes the decision into meaningful terms related to input variables. By breaking down complex decisions into interpretable components, this method enhances our understanding of how deep learning models arrive at their predictions. To evaluate these approaches, three classification tasks are performed. The results demonstrate that both methods contribute to improving interpretability by providing insights into how deep learning models make decisions. In conclusion, this paper highlights the importance of interpretability in artificial intelligence and presents two approaches for explaining predictions made by deep learning models.