Research map / Probabilistic and scientific modeling / Variational and generative sampling
$\alpha$TC-VAE: On the relationship between Disentanglement and Diversity
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 α-TCVAE, a VAE trained with a convex total-correlation lower bound combining VIB and CEB terms, and shows disentanglement correlates with generative diversity (Vendi) across datasets, with gains on MPI3D-Real and an RL Ant Maze task.
Problem
Disentangled representations are often limited to simple datasets because latent variables struggle to be informative about complex attributes; the link between disentanglement and generative diversity has not been quantitatively established.
Approach
A novel convex TC lower bound that combines the VIB and CEB terms, generalizing the beta-VAE bound (alpha=0 recovers beta-VAE), so the model maximizes both disentanglement and per-latent informativeness without extra networks or expensive sampling.
Key results
- alpha-TCVAE gives higher Vendi (more diverse) and lower FID than baseline VAEs across datasets, on par with StyleGAN in Vendi and only slightly worse in FID.
- Vendi and DCI are positively correlated across models and datasets, supporting the link between disentanglement and diversity.
- Replacing Director's beta-VAE Goal VAE with alpha-TCVAE (Alpha-Director) significantly raises mean return on the Loconav Ant Maze task, averaged over three seeds.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the α-TCVAE objective instead of baseline VAE objectives | VAE models · Vendi score across evaluated datasets | Vendi score | Helped |
| Use the α-TCVAE objective instead of baseline VAE objectives | VAE models · FID across evaluated datasets | FID | Helped |
| Use the α-TCVAE objective instead of baseline VAE objectives | VAE models · Factorized datasets | DCI | Helped |
| Use the α-TCVAE objective instead of baseline VAE objectives | VAE models · Cars3D and MPI3D-Real | SNC | Helped |
| Use α-TCVAE instead of baseline models | Attribute classifiers trained on encoded representations · Attribute classification across evaluated datasets | accuracy | Mixed |
| Replace the Director Goal VAE's β-VAE objective with the α-TCVAE bound | Director · Ant Maze | mean return | Helped |
| Replace the Director Goal VAE's β-VAE objective with the α-TCVAE bound | Director · Ant Maze sampled goals | Vendi score | Helped |
| Replace the Director Goal VAE's β-VAE objective with the α-TCVAE bound | Director · Ant Maze | exploration | Helped |
| Use a high-dimensional latent representation for CelebA | VAE models · CelebA | SNC | No clear effect |
| Sample by latent traversals rather than decoding samples from a standard normal prior | VAE-based generative models · Image generation across evaluated datasets | Vendi score | Helped |
| Sample by latent traversals rather than decoding samples from a standard normal prior | VAE-based generative models · Image generation across evaluated datasets | FID | Helped |
| Use disentangled VAEs rather than the vanilla VAE | VAE-based generative models · Image generation | Vendi score | Helped |
Surprising findings
- No model learned disentangled representations on CelebA; SNC gave all models zero, and DCI/unfairness results were judged unrealistic in high dimensions.
- Sampling from white noise gives poor diversity and reconstruction for all VAEs, while traversal-based sampling yields markedly better images.
Limitations
- Like other disentangled VAEs, alpha-TCVAE is difficult to scale to high-dimensional data; no disentanglement was achieved on CelebA.
- DCI and SNC become unreliable in high-dimensional latent spaces, and DCI may overestimate disentanglement.
Experimental setup
Datasets: Teapots 3DShapes MPI3D-Real Cars3D CelebA
Benchmarks: Teapots 3DShapes MPI3D-Real Cars3D CelebA Loconav Ant Maze
Metrics: Vendi score FID DCI SNC unfairness attribute classification accuracy mean return
Baselines: beta-VAE beta-TCVAE beta-VAE+HFS FactorVAE vanilla VAE StyleGAN Director
Builds on: beta-VAE FactorVAE beta-TCVAE variational information bottleneck conditional entropy bottleneck Director
Closest papers by meaning
- Isolating Sources of Disentanglement in Variational Autoencoders (NeurIPS 2018 · similarity 0.91)
- Disentanglement Analysis with Partial Information Decomposition (ICLR 2022 · similarity 0.90)
- An Identifiable Double VAE For Disentangled Representations (ICML 2021 · similarity 0.90)
- Disentanglement via Latent Quantization (NeurIPS 2023 · similarity 0.90)
- Synergy Between Sufficient Changes and Sparse Mixing Procedure for Disentangled Representation Learning (ICLR 2025 · similarity 0.89)
- Disentangled Representation Learning in Non-Markovian Causal Systems (NeurIPS 2024 · similarity 0.88)
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.