Research map / Probabilistic and scientific modeling / Variational and generative sampling

CASL-VAE: Learning Structured Latent Variables from Unpaired Data for Semi-supervised Clustering and Paired Sample Generation

NeurIPS 2026 · Probabilistic and scientific modeling › Variational and generative sampling · approach: Structured latent VAEs · 0 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 CASL-VAE, a contrastive VAE with a common continuous latent and a hierarchical salient latent (discrete subtypes plus continuous within-subtype factors) that learns from unpaired reference-target data for subtype clustering and paired-sample generation.

Problem

Quantifying target-population variation relative to a reference when paired data are unavailable and target variation is heterogeneous (multiple subtypes), so existing contrastive analysis methods fail to separate multiple modes of target-specific variation.

Approach

Factorizes latents into a shared continuous common latent and a hierarchical salient latent (Gumbel-Softmax discrete subtype plus conditional continuous factor), uses posterior decoupling to approximate the paired joint likelihood from unpaired data, and adds MMD, cross-domain consistency, a semi-supervised BCE on domain labels, and a KL penalty that suppresses reference information in the salient

Key results

• On semi-synthetic 10-20% atrophy, CASL-VAE achieves target-only ARI 0.468 vs 0.251 for SepVAE and 0.183 for cVAE/Smile-GAN.
• In target+reference clustering at 10-30% atrophy, CASL-VAE reaches ARI 0.697 vs 0.451 for SepVAE.
• Paired-sample generation cosine similarity is highest for CASL-VAE (0.5835 at 10-20%, 0.5816 at 10-30%) versus cVAE and SepVAE.

Surprising findings

• Under typical 10-30% atrophy, CASL-VAE is on par with Smile-GAN and K-means in target-only ARI rather than clearly best.

Limitations

• Evaluation mainly on semi-synthetic simulated atrophy; real-data validation is limited to ADNI with indirect clinical associations.
• Common latent assumed to fully encode reference variation; posterior decoupling is enforced only via regularization, not guaranteed.
• Gaussian likelihood and only 2-D continuous salient latent used; generalization to other data types untested.

Experimental setup

Datasets: iSTAGING consortium T1w MRI brain volumes ADNI

Benchmarks: semi-synthetic iSTAGING atrophy simulation (10-20%, 10-30%) ADNI1 held-out

Metrics: ARI MAE cosine similarity

Baselines: K-means GMM GMVAE Smile-GAN cVAE SepVAE

Builds on: contrastive analysis VAE Gumbel-Softmax SepVAE Smile-GAN

Closest papers by meaning

• Sparse Multi-Channel Variational Autoencoder for the Joint Analysis of Heterogeneous Data (ICML 2019 · similarity 0.88)
• Modelling Cellular Perturbations with the Sparse Additive Mechanism Shift Variational Autoencoder (NeurIPS 2023 · similarity 0.87)
• NestedVAE: Isolating Common Factors via Weak Supervision (CVPR 2020 · similarity 0.86)
• Multi-modal Gaussian Process Variational Autoencoders for Neural and Behavioral Data (ICLR 2024 · similarity 0.85)
• Censor Dependent Variational Inference (ICML 2025 · similarity 0.84)
• TreeVI: Reparameterizable Tree-structured Variational Inference for Instance-level Correlation Capturing (NeurIPS 2024 · similarity 0.84)