Modern AI techniques –- especially deep learning –- provide, in many cases, very good recommendations: where a self-driving car should go, whether to give a company a loan, etc. The problem is that not all these recommendations are good -- and since deep learning provides no explanations, we cannot tell which recommendations are good. It is therefore desirable to provide natural-language explanation of the numerical AI recommendations. The need to connect natural language rules and numerical decisions is known since 1960s, when the need emerged to incorporate expert knowledge -- described by imprecise words like "small" -- into control and decision making. For this incorporation, a special "fuzzy" technique was invented, that led to many successful applications. This book described how this technique can help to make AI more explainable.The book can be recommended for students, researchers, and practitioners interested in explainable AI.
Author(s): Vladik Kreinovich
Series: Studies in Computational Intelligence, 1047
Publisher: Springer
Year: 2022
Language: English
Pages: 135
City: Cham
Preface
Contents
1 Why Explainable AI? Why Fuzzy Explainable AI? What Is Fuzzy?
1.1 Why Explainable AI?
1.2 Why Fuzzy Techniques Seem a Reasonable Approach for Explainable AI
1.3 What Is Fuzzy Methodology
1.4 Summary of Fuzzy Methodology
1.5 Exercises
2 Defuzzification
2.1 Formulation of the Problem: Reminder
2.2 Main Idea and the Resulting Formula
2.3 Integral Form
2.4 Important Comment: Centroid Defuzzification Is Not a Panacea
2.5 Exercises
2.6 Self-Test 1
3 Which Fuzzy Techniques?
3.1 What We Study in This Chapter
3.2 Interpolation Should Be Robust
3.3 Which Interpolation Is the Most Robust
3.4 ``And''- and ``Or''-Operations Must Be Robust Too
3.5 Which Is the Most Robust ``And''-Operation
3.6 Which Is the Most Robust ``Or''-Operation
3.7 Group Robustness Versus Individual Robustness
3.8 Which Interpolation Is the Most Individually Robust
3.9 The Most Individually Robust ``And''-Operation
3.10 Robustness Versus Individual Robustness: Example
3.11 The Most Individually Robust ``Or''-Operation
3.12 General Conclusion
3.13 Exercises
4 So How Can We Design Explainable Fuzzy AI: Ideas
4.1 Machine Learning Revisited
4.2 Exercises
4.3 Self-Test 2
5 How to Make Machine Learning Itself More Explainable
5.1 How Can We Make Machine Learning Itself More Explainable: Idea
5.2 Selection of an Activation Function
5.3 Selection of Pooling
5.4 What About Fuzzy?
5.5 Exercises
5.6 Self-Test 3
6 Final Self-Test
Appendix A Terms Used in the Book (in Alphabetic Order)
Appendix B Why Do We Need …? (in Alphabetic Order)
Appendix C Solutions to Exercises
Appendix D Solutions to Self-Tests
D.1 Solutions to Self-Test 1
D.2 Solutions to Self-Test 2
D.3 Solutions to Self-Test 3
D.4 Solutions to Final Self-Test
Appendix E Additional Readings
