Research map / Probabilistic and scientific modeling / Variational and generative sampling
Var-JEPA: A Variational Formulation of the Joint-Embedding Predictive Architecture – Bridging Predictive and Generative Self-Supervised Learning
Moritz Gögl, Christopher Yau
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 |
|---|---|---|---|
| replace deterministic T-JEPA with variational Var-T-JEPA | MLP on Covertype · Covertype | test accuracy: 0.408 -> 0.679 | Improved |
| replace deterministic T-JEPA with variational Var-T-JEPA | MLP on MNIST · MNIST (tabular) | test accuracy: 0.113 -> 0.822 | Improved |
| add Var-T-JEPA embeddings to raw features baseline | MLP on Adult · Adult | test accuracy: 0.849 -> 0.852 | Improved |
| use Var-T-JEPA embeddings instead of raw features | MLP on Covertype · Covertype | test accuracy: 0.750 -> 0.679 | Hurt |
| discard 50% most uncertain samples via latent uncertainty | MLP + Var-T-JEPA · Adult | test accuracy: 0.852 -> 0.921 | Improved |
| add KL on s_x to ELBO (vs removing it) (ablation) | Var-JEPA ELBO · Simulation s_x | aggregated KL to N(0,I): 8.374 -> 0.113 | Improved |
| add KL terms on all latents (vs removing all) (ablation) | Var-JEPA ELBO · Simulation s_x | aggregated KL to N(0,I): 10.465 -> 0.113 | Improved |
| add reconstruction and generation terms (ablation) | Var-JEPA ELBO · Simulation s_x | linear-probe accuracy: 0.571 -> 0.996 | Improved |
| add SIGReg to ELBO (ablation) | Var-JEPA · Simulation s_x | aggregated KL to N(0,I): 0.113 -> 0.107 | No clear effect |
| add SIGReg to ELBO without reconstruction and generation (ablation) | ELBO (G) · Simulation s_x | linear-probe accuracy: 0.571 -> 0.834 | Improved |
| add SIGReg to no-KL ELBO (ablation) | ELBO without KL (I) · Simulation s_x | aggregated KL to N(0,I): 10.465 -> 2.115 | Improved |
| add KL on s_y to ELBO (ablation) | Var-JEPA ELBO · Simulation s_y | linear-probe accuracy: 0.983 -> 0.993 | Improved |
Closest papers by meaning
- Denoising with a Joint-Embedding Predictive Architecture (ICLR 2025 · similarity 0.90)
- How JEPA Avoids Noisy Features: The Implicit Bias of Deep Linear Self Distillation Networks (NeurIPS 2024 · similarity 0.90)
- Connecting Joint-Embedding Predictive Architecture with Contrastive Self-supervised Learning (NeurIPS 2024 · similarity 0.89)
- seq-JEPA: Autoregressive Predictive Learning of Invariant-Equivariant World Models (NeurIPS 2025 · similarity 0.89)
- Drive vs. Decay: On the Training Dynamics of Joint-Embedding Predictive Architectures (NeurIPS 2026 · similarity 0.89)
- LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels (NeurIPS 2026 · 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.
