Bernhard‘s Talk on Towards Causal NLP 笔记

因果学习系列笔记

这是我的 GitHub 因果学习笔记仓库
https://github.com/xin007-kong/ryCausalLearning，欢迎 star🤩

• 讲者是 Bernhard Schölkopf
  • talk 链接：(41) Bernhard Schoelkopf | Towards Causal NLP | Keynote@EMNLP 2021 Causal Inference and NLP Workshop - YouTube
  • 谷歌学术主页：‪Bernhard Schölkopf‬ - ‪Google Scholar‬
• 概括一下 talk 的内容
• Machine learning relies on correlations rather than causality, which can lead to limitations in object recognition and adversarial vulnerability.
• Causal inference connects correlation and causality, suggesting that statistically dependent variables have a causal explanation.

• Structural causal models provide a formalism for thinking about causality, using directed acyclic graphs (DAGs) to represent causal relationships.
• Causal factorization allows the joint distribution to be expressed as a product of conditionals, reducing the dimensionality of the variables.
• The goal is to move towards causal representation learning, combining interventional knowledge and learned representations for richer and more data-efficient world models.