Research map / Probabilistic and scientific modeling / Causal discovery

Neural Causal Abstractions

AAAI 2024 · Probabilistic and scientific modeling › Causal discovery · approach: Causal representation identifiability · 3 citations (OpenAlex)

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.

Introduces causal abstractions built from intervariable and intravariable clusters of variables and their domains, shows they can be learned with Neural Causal Models (GAN-NCM) for identification, estimation and sampling across Pearl's causal hierarchy, and extends to representation-based abstractions for image data.

Problem

Causal inference in real-world settings where the true low-level SCM (e.g., pixel-level) is unknown and only limited low-level data is available, so high-level causal abstractions must be learned rather than declared.

Approach

Defines constructive abstraction functions from inter-variable clusters (grouping variables) and intra-variable clusters (partitioning value domains), a layer-wise (PCH) consistency notion, and the abstract invariance condition; then learns high-level SCMs as G_C-constrained neural causal models over the cluster-level variables, reducing abstract identification to neural identification.

Key results

• Worked nutrition/drug examples show a non-identifiable query under one clustering (values 0.95 vs 0.29) becomes identifiable under an expanded clustering, matching the true 0.5.
• Proposes NeuralAbstractID procedure that is sound and complete for abstract identification given L2 data and a C-DAG (corollary, theoretical).
• Experiments on high-dimensional image data are reported as supporting the theory (details beyond the extracted text not shown).

Surprising findings

• Abstract causal hierarchy theorem: a high-level model matching the lower layer (L1) will almost never match higher layers (L2, L3) under abstraction.
• Naively grouping values by sum (total cholesterol) violates the abstract invariance condition, while grouping by difference satisfies it.

Limitations

• Theory assumes recursive SCMs with finite discrete domains and admissible clusters given in advance.
• Identification requires a correct C-DAG over high-level variables; unobserved confounding can make queries non-identifiable.

Experimental setup

Datasets: simulated SCM data image data

Metrics: interventional and counterfactual query values vs ground tru

Builds on: Neural Causal Models causal abstraction theory (Beckers and Halpern) NeuralID GAN-NCM C-DAG

Closest papers by meaning

• Causal Abstraction Inference under Lossy Representations (ICML 2025 · similarity 0.92)
• The Non-Linear Representation Dilemma: Is Causal Abstraction Enough for Mechanistic Interpretability? (NeurIPS 2025 · similarity 0.91)
• Learning Linear Causal Representations from General Environments: Identifiability and Intrinsic Ambiguity (NeurIPS 2024 · similarity 0.89)
• A Sparsity Principle for Partially Observable Causal Representation Learning (ICML 2024 · similarity 0.88)
• Learning Discrete Concepts in Latent Hierarchical Models (NeurIPS 2024 · similarity 0.88)
• Causal Abstraction Learning based on the Semantic Embedding Principle (ICML 2025 · similarity 0.88)