2. Before XAI
2
• AI tech(ML) provides decision making, recommendation, prediction based on big
data and/or complicate algorithm
• Some ML(deep learning) is called “Black box” because of complicate algorithm
• There are some issues that it cannot provide reason of derived result, validity of derivation
process
• However, at AI system that treats personal info & assets based on customer’s trust
• ex) finance, insurance, medical field
• It needs verification about reason of derived result, validity of derivation process
• to assure fairness, reliability, accuracy
4. background of appearance
4
• Fears of user & society’s acceptance and confidence about AI system
Increase of interest in XAI
• 1970’s AI system cannot understanding derived results to experts
• Importance of explainable AI is recognized
• research started by few researchers
• Recently, deep learning spreads worldwide and adopts various fields
XAI rises again
• DARPA(The Defense Advanced Research Projects Agency) starts XAI project(XAI
learning model develop & test) since ‘17
5. Concept
5
• Technique that User understands AI system’s operation & results, interprets
correctly, be explainable to generate results
• Ex) Classifying cat images
• Traditional system: derives only either cat or not
• XAI: derives cat or not, provides reason(hair, whiskers, etc) to user
6. Technical approach for XAI
6
❖ Deform existing learning model
❖ Develop new learning model
❖ Compare between learning models
7. Deform existing learning model
7
• Existing model
• add inverse calculation(ex. Back prop.)
• Edit learning model
9. Develop new learning model
9
• Reason-result process is Explainable
• Ex) Learning AND-OR Templates for Object Recognition and Detection
10. Compare between learning models
10
• Explain final result by comparing other explainable model without specific
information about learning model trying to explain
• Ex) classification model that compares pixels with a descriptive learning model for
the description of the image classification model
11. Compare between learning models
11
- When XAI classifies Figure 6(a) as an electronic guitar, the pixels used in XAI (the
partial value of the neck of the electric guitar) are applied to target model
- the basis of XAI can be used as the basis for target model's results
- if similar results are obtained
19. Conclusion
19
• AI is recognized as a key technology in future
• But some people worry about discrimination and inequality in society
• Due to using bigdata like artificial intelligence
• In these concenrs, XAI expects AI in various areas(financial, insurance, etc.)
• To gain trust from users and customers
• To create a consensus for social acceptance
• Technical approach to XAI
• transformed to explain existing learning models
• new learning models can be developed
• methods by comparing learning models exist.
20. Conclusion
20
• the effects By utilizing XAI
• improving the performance of the artificial intelligence system
• Identify system performance degradation factors, such as learning model bias
• derive appropriate learning models from comparisons
• between learning models of the same purpose and result
• acquiring surveillance skills
• Extracts and analyzes various patterns from big data
• during the learning process to derive laws, strategies, etc.
• verifying legal responsibility and compliance
• Incorrect results of artificial intelligence systems can cause conflicts, and can be verified for
compliance with regulations
• Ex) the European Union's Privacy Regulations (GDPR)
21. References
21
• [설명 가능한 인공지능(eXplainable AI, XAI) - 금융보안원]
(https://www.fsec.or.kr/common/proc/fsec/bbs/42/fileDownLoad/1447.do)
[<한국경제> 에이젠글로벌, 인공지능 `설명력(XAI)` 금융특화 모델 제시]
(https://www.hankyung.com/news/article/2019081921305)
[<venturebeat> Fiddler raises $10.2 million for AI that explains its reasoning]
(https://venturebeat.com/2019/09/24/fiddler-raises-10-2-million-for-ai-that-explains-
its-reasoning/)
22. References - book
22
• [Explainable AI: Interpreting, Explaining and Visualizing Deep Learning]
(https://books.google.co.kr/books?id=j5yuDwAAQBAJ&pg=PA282&dq=explainable
+ai&hl=en&sa=X&ved=0ahUKEwiC-aiD9KrlAhUIHXAKHVR7CJwQ6AEIKDAA)
• [Explainable and Interpretable Models in Computer Vision and Machine
Learning](https://books.google.co.kr/books?id=oKR8DwAAQBAJ&pg=PA19&dq=ex
plainable+ai&hl=en&sa=X&ved=0ahUKEwiC-
aiD9KrlAhUIHXAKHVR7CJwQ6AEIPzAE)
• [Reasoning Web: Explainable Artificial Intelligence]
(https://books.google.co.kr/books?id=PkewDwAAQBAJ&pg=PA277&dq=explainabl
e+ai&hl=en&sa=X&ved=0ahUKEwiC-aiD9KrlAhUIHXAKHVR7CJwQ6AEIRDAF)