Research map / Probabilistic and scientific modeling / Variational and generative sampling
A Probabilistic Hard Concept Bottleneck for Steerable Generative Models
María Martínez-García, Ricardo Vazquez Alvarez, Alejandro Lancho, Pablo M. Olmos, Isabel Valera
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use VHCB hard probabilistic concept bottleneck instead of CB-AE | StyleGAN2 CelebA-HQ · single intervention inactive to active, low corr., ResNet18 | target accuracy: 0.42 -> 0.77 | Improved |
| use VHCB instead of CB-AE | StyleGAN2 CelebA-HQ · single intervention, non-target | non-target accuracy: 0.83 -> 0.76 | Hurt |
| use VHCB instead of CB-AE | StyleGAN2 CelebA-HQ · concept inference, all concepts, CLIP labels | accuracy: 0.565 -> 0.623 | Improved |
| use VHCB instead of CB-AE | StyleGAN2 CelebA-HQ · disentanglement of c and s, low corr., ResNet18 | accuracy: 0.701 -> 0.874 | Improved |
| sample concept patterns by training frequency instead of random concept sets | VHCB · direct generation, all concepts, ResNet18 | target accuracy: 0.551 -> 0.873 | Improved |
| target minimum Hamming distance training patterns instead of single arbitrary concept | VHCB · interventions, all concepts, ResNet18, active to inactive | target accuracy: 0.550 -> 0.660 | Improved |
| use a low-correlation concept set instead of all 40 concepts | VHCB · single intervention inactive to active, ResNet18 | target accuracy: 0.170 -> 0.769 | Improved |
| add unsupervised side channel s (ablation) | VHCB · CelebA-HQ balanced concepts, FID | FID: 20.950 -> 11.016 | Improved |
| add unsupervised side channel s (ablation) | VHCB · CelebA-HQ all concepts, FID | FID: 7.921 -> 7.248 | No clear effect |
| use symmetric KL instead of BCE concept loss (ablation) | VHCB · CelebA-HQ balanced, ResNet18, a to i | target accuracy: 0.564 -> 0.814 | Mixed |
| use symmetric KL instead of BCE concept loss (ablation) | VHCB · CelebA-HQ balanced, ResNet18, i to a | target accuracy: 0.395 -> 0.375 | Hurt |
| increase side channel KL weight from 0 to 40 (ablation) | VHCB · CelebA-HQ balanced, s size 5, generation | target accuracy: 0.659 -> 0.561 | Hurt |
Closest papers by meaning
- Interpretable Generative Models through Post-hoc Concept Bottlenecks (CVPR 2025 · similarity 0.91)
- What's Producible May Not Be Reachable: Measuring the Steerability of Generative Models (NeurIPS 2025 · similarity 0.89)
- Interpretable but Fragile? Robustness of Concept Bottlenecks under Geometric-Semantic Perturbations (NeurIPS 2026 · similarity 0.89)
- Hyperbolic Concept Bottleneck Models (NeurIPS 2026 · similarity 0.88)
- Hybrid Concept Bottleneck Models (CVPR 2025 · similarity 0.88)
- Stochastic Concept Bottleneck Models (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.
