Research map / Probabilistic and scientific modeling / Causal discovery
CaRiNG: Learning Temporal Causal Representation under Non-Invertible Generation Process
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 CaRiNG, an identifiability theory and sequential-VAE-based method that recovers time-delayed latent causal variables from non-invertible temporal generation by using temporal context in the encoder and a normalizing-flow transition prior.
Problem
Identifying time-delayed latent causal processes from sequential observations when the mixing function from latents to observations is non-invertible (information loss such as occlusion or persistence of vision), where existing nonlinear ICA methods assume invertibility.
Approach
Use a sequence-to-step encoder that maps a window of observations x_{t:t-mu} to latent z_t so lost information is recovered from temporal context, combined with a normalizing-flow transition prior enforcing conditional independence of latents; theory shows identifiability up to permutation and component-wise transformation even under non-invertible mixing.
Key results
- On synthetic NG setting, CaRiNG reaches MCC 0.933 versus 0.627 for TDRL and 0.750 for Sequential VAE.
- On NG-TDMP setting, CaRiNG reaches MCC 0.921 versus 0.847 for Sequential VAE and 0.837 for TDRL.
- On SUTD-TrafficQA, CaRiNG achieves 41.22% accuracy, versus 38.58% for CMCIR and 37.32% for TDRL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CaRiNG instead of SequentialVAE | Temporal causal representation learning · NG | MCC: 0.750 \textpm 0.035 -> 0.933 \textpm 0.010 | Helped |
| Use CaRiNG instead of SequentialVAE | Temporal causal representation learning · NG-TDMP | MCC: 0.847 \textpm 0.019 -> 0.921 \textpm 0.010 | Helped |
| Use CaRiNG with causal representation learning instead of HCRN alone | HCRN on SUTD-TrafficQA · SUTD-TrafficQA | accuracy: 36.26 -> 41.22 | Helped |
| Use CaRiNG instead of CMCIR | VideoQA on SUTD-TrafficQA · SUTD-TrafficQA | accuracy: 38.58 -> 41.22 | Helped |
| Add temporal context to the encoder | Temporal representation learning under non-invertible generation · NG and NG-TDMP | MCC | Helped |
| Add the conditional-independence KL constraint | Sequential VAE-based representation learning · NG and NG-TDMP | MCC | Helped |
Surprising findings
- TDRL, which performs well under invertible generation, degrades markedly as non-invertibility increases.
- The paper argues that the commonly used Jacobian condition (at most one nonzero entry per row) alone is insufficient for identifiability in nonlinear settings, requiring path-connectedness and non-degeneracy assumptions.
Limitations
- Assumes no instantaneous causal relations, only time-delayed influences.
- Assumes the number of latent variables stays constant across time steps (no entering or exiting agents).
- Real-world datasets lack ground-truth latent variables, so identifiability is only validated on synthetic data.
Experimental setup
Datasets: SUTD-TrafficQA synthetic non-invertible temporal simulation data (NG, NG-TDMP) Volleyball
Benchmarks: SUTD-TrafficQA Volleyball synthetic NG and NG-TDMP settings
Metrics: Mean Correlation Coefficient (MCC) accuracy
Baselines: TDRL LEAP SlowVAE PCL betaVAE SKD iVAE Sequential VAE I3D+LSTM HCRN VQAC MASN
Builds on: TDRL iVAE Sequential VAE LEAP HCRN
Closest papers by meaning
- Towards Identifiability of Hierarchical Temporal Causal Representation Learning (NeurIPS 2025 · similarity 0.90)
- Causal Temporal Representation Learning with Nonstationary Sparse Transition (NeurIPS 2024 · similarity 0.90)
- Causal Representation Learning Made Identifiable by Grouping of Observational Variables (ICML 2024 · similarity 0.90)
- Learning Linear Causal Representations from General Environments: Identifiability and Intrinsic Ambiguity (NeurIPS 2024 · similarity 0.89)
- Multi-View Causal Representation Learning with Partial Observability (ICLR 2024 · similarity 0.89)
- A Sparsity Principle for Partially Observable Causal Representation Learning (ICML 2024 · similarity 0.89)
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.