When Do Learned Blocks Preserve Causal Meaning? Block Identifiability and Constructive Causal Abstractions
Standard identifiability theory assumes one-dimensional latent factors, but learned representations can encode latent information across multidimensional blocks whose internal coordinates are not uniquely determined.
ProofPaper ↗
Key points
- When such blocks are used to represent and intervene on higher-level concepts, it is unclear whether their causal meaning survives the transformations left unresolved by block-identifiability.
- We connect block-identifiability to constructive causal abstraction.
- We show that block-diagonal invertible transformations followed by compatible coordinate permutations form the maximal class that preserves every constructive causal abstraction built on those blocks.
- A small synthetic experiment with additive autoencoders shows that high-level interventions can be transferred between independently learned representations while approximately preserving their intended effects.
Sources (1)
- [1]When Do Learned Blocks Preserve Causal Meaning? Block Identifiability and Constructive Causal AbstractionsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 11:00 AM
Standard identifiability theory assumes one-dimensional latent factors, but learned representations can encode latent information across multidimensional blocks whose internal coordinates are not uniquely determined.
When such blocks are used to represent and intervene on higher-level concepts, it is unclear whether their causal meaning survives the transformations left unresolved by block-identifiability.
Extractive summary: sentences quoted from the sources.