AI and Society(1)

1. A realistic view of AI

Goldilocks rule for AI

• Toop optimistic: Sentient / super-intelligent AI killer robots coming soon
• Too pessimistic: AI cannot do everything, so an AI winter is coming
• Just right: AI can't do everything, but will transform industries

 

Limitations of AI

• Performance limitations
• Explainability is hard (but sometimes doable)
• Biased AI through biased data
• Adversarial attacks on AI

 

 

2. Discrimination / Bias

AI learning unhealthy stereotypes

• Man : Woman as Father : Mother
• Man : Woman as King : Queen
• Man : Computer programmer as Woman : Homemaker → Computer programmer

man과 computer programmer 사이의 관계를 통해 woman과 homemaker의 관계를 유추하는 부적절한 상황

 

Why bais matters

• Hiring tool that discriminated against women
• Facial recognition working better for light-skinned than dark-skinned individuals
• Bank loan approvals

 

Combating bias

• Techinical solutions:
  • E.G., "zero out" the bias in words
  • Use less biased and/or more inclusive data
• Transparency and/or auditing processes
• Diverse workforce
  • Creates less biased applications

 

 

3. Adversarial attacks on AI

Adversarial attacks on AI

• 이미지를 분류함에 있어서 약간의 변화를 주더라도 컴퓨터는 이를 구분할 방법이 없다.

Physical attacks

• 예를 들어 stop 사인을 stop으로 인지하지 못하게 만들 수도 있다.
  그런데 이런 과정이 아주 작은 스티커 하나만을 더함으로써 진행될 수 있으므로 의도적으로 학습을 방해하기가 아주 쉽다는 것이다.

 

Adversarial defenses

• Defenses do exist, but incur some cost
• Similar to spam vs. anti-spam, we may be in an arms race for some applications

 

 

 

출처: Coursera, AI For Everyone, DeepLearning.AI