Research map / Probabilistic and scientific modeling / Causal discovery
Learning Independent Causal Mechanisms
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply the trained expert that wins each transformed input before classification | Competing experts · MNIST transformed test digits | classifier accuracy: 40% accuracy ... almost reached the upper bound | Helped |
| Use approximate identity initialization instead of random initialization (ablation) | Competing experts · MNIST transformed data, 10 mechanisms | successful runs: $7$ out of the $10$ runs -> only one experiment | Helped |
| Reduce canonical-distribution training examples from 30,000 to 64 (ablation) | Competing experts · MNIST transformed test digits | classifier accuracy: 30\,000 down to 64; 96% | Hurt |
| Use 6 experts rather than 10 for 10 mechanisms (ablation) | Competing experts · MNIST transformed test digits | classifier accuracy: does not exceed 72% | Hurt |
| Use 16 experts rather than 10 for 10 mechanisms (ablation) | Competing experts · MNIST transformed test digits | classifier accuracy | No clear effect |
| Replace the committee with one network using 32, 64, or 128 filters per layer (ablation) | Competing experts · MNIST inverse transformations | number of inverse mechanisms learned: none ... correctly learn more than one inverse mechanism | Hurt |
| Increase expert width from 32 to 64 or 128 filters per layer (ablation) | Competing experts · MNIST inverse transformations | specialization: 32 ... to 64 or 128 | No clear effect |
| Increase receptive field with two pooling and two upsampling layers (ablation) | Competing experts · MNIST inverse transformations | specialization | No clear effect |
| Apply learned inverse mechanisms to Omniglot characters not seen in training | Competing experts trained on MNIST · Omniglot transformed characters | mechanism generalization | Helped |
| Compose experts for noise, up-left translation, and contrast inversion | Competing experts trained on MNIST · Omniglot characters with three consecutive transformations | reconstruction | Helped |
Closest papers by meaning
- A Fixed-Point Approach for Causal Generative Modeling (ICML 2024 · similarity 0.88)
- Identifiable Exchangeable Mechanisms for Causal Structure and Representation Learning (ICLR 2025 · similarity 0.87)
- Learning Structural Causal Models from Ordering: Identifiable Flow Models (AAAI 2025 · similarity 0.86)
- Learning Linear Causal Representations from General Environments: Identifiability and Intrinsic Ambiguity (NeurIPS 2024 · similarity 0.86)
- Toward Causal Generative Modeling: From Representation to Generation (AAAI 2025 · similarity 0.85)
- Identifying General Mechanism Shifts in Linear Causal Representations (NeurIPS 2024 · similarity 0.85)
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.