Research map / Probabilistic and scientific modeling / Causal discovery
Representation Learning: A Causal Perspective
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a causal framework that formalizes non-spuriousness, efficiency, and disentanglement in representation learning using probabilities of sufficiency and necessity and causal identification.
Problem
Desirable properties of learned representations (non-spuriousness, efficiency, disentanglement) are intuitive but hard to define formally, measure, and optimize.
Approach
Uses causal notions to define non-spuriousness and efficiency in supervised representation learning via probabilities of sufficiency and necessity, and defines disentanglement as absence of causal connections among features in unsupervised learning, with causal identification studied for observability.
Key results
- Defines non-spuriousness and efficiency for supervised representations using probabilities of sufficiency (PS) and necessity (PN).
- Formalizes disentanglement in unsupervised settings as absence of causal connections among learned features.
- Studies observable implications of these causal criteria via causal identification.
Limitations
- Causal quantities such as PS and PN are not always identifiable from observational data.
- Excerpt provides no empirical validation of the proposed metrics.
Experimental setup
Builds on: probabilities of sufficiency and necessity
Closest papers by meaning
- Learning Linear Causal Representations from General Environments: Identifiability and Intrinsic Ambiguity (NeurIPS 2024 · similarity 0.88)
- Disentangled Representation Learning in Non-Markovian Causal Systems (NeurIPS 2024 · similarity 0.87)
- Causal Representation Learning Made Identifiable by Grouping of Observational Variables (ICML 2024 · similarity 0.87)
- Multi-View Causal Representation Learning with Partial Observability (ICLR 2024 · similarity 0.87)
- A Sparsity Principle for Partially Observable Causal Representation Learning (ICML 2024 · similarity 0.87)
- Object centric architectures enable efficient causal representation learning (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.