[XAI Reading Log] Layer-Wise Relevance Propagation #1

About XAI Reading Log:  학교 영어시간 중 영어 책을 읽고 리딩로그를 작성하는 수행평가를 1학기 동안 진행한다고 한다. 이 시간에 XAI 분야를 공부하겠다고 마음먹고 2주마다 1.5개의 논문을 훝는 것을 목표를 두었다. 별도의 리딩로그 양식을 제작하여 읽기로 영어쌤과 얘기를 나눴고...그래서 매주 1개의 리딩로그를 올릴 예정이다! 이 기회를 통해 논문과 수식에 익숙해지고 싶다.  현실은 영어로 간단한 문장 하나 만들기도 버거워하지만 말이다...

---

 

Date: 2021/08/26 

Title: On pixel-wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation #1

 

- Today’s Raeding Purpose: Before reading the paper in detail, I would like to summarize the overall concept and figure out what I want to read in detail in the next time.

 

- Reasons why I want to read this: Recently, I got interested in the XAI(eXplainable AI) field and bought a book called "XAI: 설명 가능한 인공지능, 인공지능을 해부하다." In the neural network visualization part, there was interesting theory that can visuallize heatmaps of AI models, called Layer-wise relevance propagation. Though LRP isn’t the latest XAI technique to output heat maps, we can easily understand basic concepts of visuallizing heatmap important to be the foundational knowlede of other papers related XAI.

 

- Pre-view concept of LRP: The LRP backtracks and decomposes the results, displaying the relative contribution of each component as a heat map. There’re two significant processs of LRP, ‘Pixel-wise Decomposition’ and ‘Relevance Propagation’. Pixel-Wise Composition is the process of decompose the contribution of pixels in an input image to a deep learning model. Decomposition is the return of a heat map by determining the contribution of each layer of the model to the results, moving backward from the possibility of predicting to a particular class. Relevance Propagation is a method of calculating the validity of how the hidden layer which has completed the decomposition process, contributes to the output of the result value.

 

- Today’s Important Formulas :

• Relevance R 

  

  

  

• Taylor Type Decomposition

  

  

 

My thoughts of today’s reading:

Today, I learned general concept of LRP and Taylor Decomposition as a way to approximately evaluate pixel-wise relevacne score. Decomposing process was difficult to understand. I will try to read this, focusing on applying practical LRP cases. I hope this opportunity will help me be used to reading Computer Science Papers.

 

Reference:

[1] 안재현,『XAI: 설명 가능한 인공지능, 인공지능을 해부하다.』, 위키북스, 209~242pg

[2] Bach, Sebastian, et al. “On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation.” PLOS ONE, edited by Oscar Deniz Suarez, vol. 10, no. 7, 2015, p. e0130140. Crossref, doi:10.1371/journal.pone.0130140.