Research map / Probabilistic and scientific modeling / Variational and generative sampling
Representation Uncertainty in Self-Supervised Learning as Variational Inference
Hiroki Nakamura, Masashi Okada, Tadahiro Taniguchi
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 |
|---|---|---|---|
| learn per-image uncertainty kappa with Power Spherical latent (VI-SimSiam) instead of SimSiam | SimSiam · ImageNet100 linear eval, 50 epochs | Top-1 accuracy: 41.07 -> 63.89 | Improved |
| VI-SimSiam instead of SimSiam at long training | SimSiam · ImageNet100 linear eval, 500 epochs | Top-1 accuracy: 81.61 -> 77.49 | Hurt |
| VI-SimSiam instead of SimSiam on CIFAR-10 | SimSiam · CIFAR-10 linear eval, 100 epochs | Top-1 accuracy: 77.43 -> 89.92 | Improved |
| use predicted kappa as confidence score for correctness | VI-SimSiam (500 epochs) · ImageNet100 | AUROC: 0.5 (chance) -> 0.72 linear, 0.76 KNN | Improved |
| use negative entropy of DINO representation as confidence score | DINO (200 epochs) · ImageNet100 | AUROC: 0.5 (chance) -> 0.67 | Improved |
| random crop augmentation compared to other augmentations (ablation) | VI-SimSiam kappa · ImageNet100 val | mean kappa: 852.63 (normal) -> 812.26 (random crop) | Hurt |
| select the highest-kappa view among 30 augmentations at test time (ablation) | VI-SimSiam · ImageNet100 linear eval, 200 epochs | Top-1 accuracy: 76.74 -> 77.36 | No clear effect |
| VI-SimSiam instead of SimSiam for transfer with 100-epoch pretraining | SimSiam · Flowers transfer linear eval | Top-1 accuracy: 14.8 -> 64.5 | Improved |
| VI-SimSiam instead of SimSiam for transfer with 500-epoch pretraining | SimSiam · Food transfer linear eval | Top-1 accuracy: 59.2 -> 46.1 | Hurt |
| VI-SimSiam instead of SimSiam with 1% labels | SimSiam · ImageNet100 semi-supervised, 1% labels, 100 epochs | Top-1 accuracy: 33.8 -> 53.5 | Improved |
Closest papers by meaning
- Laplacian Autoencoders for Learning Stochastic Representations (NeurIPS 2022 · similarity 0.88)
- Remedying Uncertainty Representations in Visual Inference Through Explaining-Away Variational Autoencoders (TPAMI 2026 · similarity 0.88)
- Nonparametric Variational Auto-Encoders for Hierarchical Representation Learning (ICCV 2017 · similarity 0.88)
- Consistency Regularization for Variational Auto-Encoders (NeurIPS 2021 · similarity 0.87)
- Variational Autoencoders Pursue PCA Directions (by Accident) (CVPR 2019 · similarity 0.87)
- Maxitive Donsker-Varadhan Formulation for Possibilistic Variational Inference (NeurIPS 2026 · similarity 0.86)
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.
