Research map / Probabilistic and scientific modeling / Causal discovery

Causal Representation Learning from Multimodal Biomedical Observations

ICLR 2025 · Probabilistic and scientific modeling › Causal discovery · approach: Causal representation identifiability · 1 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.

Proposes identifiability theory for nonparametric multimodal latent causal models, showing subspace identifiability under cross-modal influence and component-wise identifiability under sparse cross-modal causal connections, implemented as a flow-based encoder-decoder with independence and L1 sparsity regularizers.

Problem

Learning interpretable, identifiable latent causal variables from multimodal biomedical observations, where existing multimodal CRL methods assume restrictive latent distributions or only yield subspace-level identification.

Approach

Nonparametric latent distributions with cross-modal causal links; subspace identifiability needs only that each modality's latents influence other modalities (local invertibility), and component-wise identifiability follows from sparsity of cross-modality latent causal edges, enforced by an L1 penalty on a learnable adjacency matrix within a flow-based encoder-decoder.

Key results

• On Variant MNIST, the proposed method achieves MCC 0.85 and R2 0.90, versus 0.74 MCC for CausalVAE, 0.63 for MCL, and 0.53 for BetaVAE.
• In numerical simulations the inter-modal causal graph is recovered with SHD=0 in the reported case.
• Identifiability improves with higher sparsity ratio of inter-modal latent causal links, and the framework outperforms baselines in all sparsity scenarios.

Measured effects reported in the paper

Surprising findings

• CausalVAE achieves a higher MCC (0.74) than the proposed method's competitors MCL and BetaVAE on Variant MNIST yet has lower R2 (0.50), indicating metric disagreement.

Limitations

• Assumes known number of latent variables per modality.
• No quantitative comparison against biomedical benchmark models.
• Real-world results are assessed qualitatively against medical literature.

Experimental setup

Datasets: numerical simulations Variant MNIST (colored MNIST + fashion MNIST) human phenotype dataset

Benchmarks: numerical simulations Variant MNIST human phenotype dataset

Metrics: MCC R2 SHD

Baselines: BetaVAE CausalVAE MCL (multimodal contrastive learning)

Builds on: multimodal CRL subspace identifiability CausalVAE BetaVAE multimodal contrastive learning

Closest papers by meaning

• Learning Linear Causal Representations from General Environments: Identifiability and Intrinsic Ambiguity (NeurIPS 2024 · similarity 0.89)
• Identifiable Latent Polynomial Causal Models through the Lens of Change (ICLR 2024 · similarity 0.88)
• Causal Representation Learning Made Identifiable by Grouping of Observational Variables (ICML 2024 · similarity 0.88)
• A Sparsity Principle for Partially Observable Causal Representation Learning (ICML 2024 · similarity 0.88)
• Propensity Score Alignment of Unpaired Multimodal Data (NeurIPS 2024 · similarity 0.87)
• Differentiable Causal Discovery for Latent Hierarchical Causal Models (ICLR 2025 · similarity 0.87)