Weakly supervised causal representation learning
What this work does
Uses paired observations before/after unknown interventions to learn latent causal variables and mechanisms, with explicit and implicit causal-model parameterizations.
Connection to discovery
A precise precedent for representation-level discovery with weak experimental supervision, rather than treating every new embedding as a discovered variable.
Evidence
Inspected the equivalence definition, smooth/invertible model assumptions and the VAE-based implementation and downstream intervention tasks.
Scope and limitations
Identifiability is up to stated equivalences under substantial assumptions. Unknown intervention targets do not mean an unconstrained or intervention-free setting.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.