- 1、本文档共35页,可阅读全部内容。
- 2、原创力文档(book118)网站文档一经付费(服务费),不意味着购买了该文档的版权,仅供个人/单位学习、研究之用,不得用于商业用途,未经授权,严禁复制、发行、汇编、翻译或者网络传播等,侵权必究。
- 3、本站所有内容均由合作方或网友上传,本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺!文档内容仅供研究参考,付费前请自行鉴别。如您付费,意味着您自己接受本站规则且自行承担风险,本站不退款、不进行额外附加服务;查看《如何避免下载的几个坑》。如果您已付费下载过本站文档,您可以点击 这里二次下载。
- 4、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等,请点击“版权申诉”(推荐),也可以打举报电话:400-050-0827(电话支持时间:9:00-18:30)。
查看更多
Learning Causality and Learning with
Causality: A Road to Intelligence
Kun Zhang
kunz 1@
Carnegie Mellon University
Categorization by Google
Photos
Google apologizes for algorithm mistakenly calling black people
'gorillas'
Adversarial Attack
(Goodfellow, 2015)
Artificial “Intelligence”
•Traditional machine learning usually assumes a fixed data
distribution; avoids overfitting
•Intelligence: understanding; control/intervention;
decomposability; information fusion, learning with few
examples, extrapolation
To Achieve “Artificial” Intelligence
• Intelligence?
• An evolution, selection, and growth perspective
• Two components
• Survival: Good prediction/decision across
scenarios
• Inner compact representation…
• Creativity
• Causal representation
• The two types of representations seem consistent
Causality vs. Association
Causality vs. Dependence
•Causality ➜ dependence ! Dependence ➜ causality
(/comics/correlation.png)
3
Outline
•Learning causality X Y
•Conditional independence X Y
Z
•Independent noise X Y
•Independent changes
•Transfer learning
•Modularity or independent changes
Reveal Causality from
Observational Data?
• Connection between causal structure and
statistical properties of the data under
suitable assumptions ?
• Properties of causal systems: modularity
If there is no common cause of X and Y,
文档评论(0)