Results of the summarizing process for the arXiv paper: 2012.09838v2

In the field of text processing and computer vision classification tasks, self-attention techniques, particularly Transformers, have gained significant prominence. These methods rely on attention maps or heuristic propagation to visualize the image components that contribute to a specific classification. However, existing approaches face challenges when it comes to attention layers and skip connections. To address these challenges, this work proposes a novel method for computing relevancy in Transformer networks. The proposed method utilizes the Deep Taylor Decomposition principle to assign local relevance scores and then propagates them through the layers. This propagation involves attention layers and skip connections, which pose difficulties for existing methods. The formulation used in this solution ensures that total relevancy is maintained across layers. To evaluate the effectiveness of the proposed method, it is benchmarked on recent visual Transformer networks as well as a text classification problem. The results demonstrate a clear advantage over existing explainability methods. In further analysis, different variants of the proposed method are compared in terms of performance. It is found that considering all attention layers leads to the best results while maintaining relevance and gradient integration. However, considering only the last attention layer still yields moderate performance. Among single block visualizations, focusing on the block closest to the output proves more informative than the one closest to the input. The authors acknowledge funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation program. The first author's contribution is part of a project conducted at Tel Aviv University. Overall, this work presents specific solutions to challenges faced by existing methods in interpreting Transformers' attentions.

• - Self-attention techniques, particularly Transformers, are prominent in text processing and computer vision classification tasks
• - Existing approaches face challenges with attention layers and skip connections
• - The proposed method uses Deep Taylor Decomposition to compute relevancy in Transformer networks
• - Total relevancy is maintained across layers in the proposed method
• - The proposed method outperforms existing explainability methods in visual Transformer networks and text classification problems
• - Considering all attention layers leads to the best results while maintaining relevance and gradient integration
• - Focusing on the block closest to the output is more informative than the one closest to the input in single block visualizations
• - Funding for this work was provided by the European Research Council (ERC) under Horizon 2020 program

Summary1. Transformers are important for understanding and classifying text and pictures. 2. There are problems with how attention is used in current methods. 3. A new method called Deep Taylor Decomposition is used to calculate importance in Transformers. 4. The new method keeps importance consistent across different layers. 5. The new method works better than other ways of explaining Transformers. Definitions- Self-attention techniques: Methods that help a computer understand and classify information by paying attention to different parts of it. - Transformers: A type of self-attention technique that is commonly used in tasks like understanding text and pictures. - Relevancy: How important something is in a given context or situation. - Total relevancy: The overall importance of something, taking into account all the relevant factors or aspects. - Explainability methods: Ways to understand and explain how a computer makes decisions or processes information.

Exploring the Relevancy of Transformer Networks

In recent years, self-attention techniques have become increasingly popular in text processing and computer vision classification tasks. These methods rely on attention maps or heuristic propagation to visualize the components that contribute to a specific classification. However, existing approaches face challenges when it comes to attention layers and skip connections. To address these issues, this paper proposes a novel method for computing relevancy in Transformer networks.

The Deep Taylor Decomposition Principle

The proposed method utilizes the Deep Taylor Decomposition principle (DTD) to assign local relevance scores and then propagates them through the layers. This propagation involves attention layers and skip connections, which pose difficulties for existing methods. The formulation used in this solution ensures that total relevancy is maintained across layers.

Evaluating Performance

To evaluate the effectiveness of the proposed method, it was benchmarked on recent visual Transformer networks as well as a text classification problem. The results demonstrate a clear advantage over existing explainability methods. In further analysis, different variants of the proposed method were compared in terms of performance. It was found that considering all attention layers leads to the best results while maintaining relevance and gradient integration; however, considering only the last attention layer still yields moderate performance among single block visualizations focusing on blocks closest to output proves more informative than those closest to input..

Acknowledgements

The authors acknowledge funding from European Research Council (ERC) under European Union's Horizon 2020 research and innovation program as well as contributions from Tel Aviv University where first author's contribution is part of project conducted there .

Conclusion

Overall, this work presents specific solutions to challenges faced by existing methods in interpreting Transformers' attentions with improved accuracy over previous models while also providing insights into how individual components are contributing towards overall result .