Research map / Label-efficient and robust learning / Knowledge distillation
Hyper-Representations as Generative Models: Sampling Unseen Neural Network Weights
Konstantin Schürholt, Boris Knyazev, Xavier Giró-i-Nieto, Damian Borth
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 |
|---|---|---|---|
| add layer-wise loss normalization to the autoencoder (ablation) | hyper-representation + KDE30 sampling · SVHN, epoch 0 | test accuracy (%): 10.1 -> 51.5 | Improved |
| add layer-wise loss normalization (ablation) | hyper-representation + KDE30 · CIFAR-10, epoch 0 | test accuracy (%): 15.5 -> 26.9 | Improved |
| add layer-wise loss normalization (ablation) | hyper-representation + KDE30 · STL-10, epoch 0 | test accuracy (%): 12.7 -> 19.7 | Improved |
| add layer-wise loss normalization (ablation) | hyper-representation + KDE30 · MNIST, epoch 0 | test accuracy (%): 63.2 -> 68.6 | Improved |
| add layer-wise loss normalization, after 25 epochs of fine-tuning (ablation) | hyper-representation + KDE30 · SVHN, epoch 25 | test accuracy (%): 75.4 -> 74.2 | No clear effect |
| initialize with sampled weights and fine-tune 25 epochs instead of training from scratch | S_KDE30 · MNIST | test accuracy (%): 83.3 -> 93.0 | Improved |
| initialize with sampled weights, 25 epochs vs scratch 50 epochs | S_KDE30 · SVHN | test accuracy (%): 70.7 -> 74.2 | Improved |
| ensemble 15 sampled models instead of single scratch-trained model (50 epochs) | S_KDE30 · SVHN | test accuracy (%): 70.7 -> 77.6 | Improved |
| use top-30% sampled population instead of 100% (ablation) | KDE sampling · MNIST | test accuracy (%) | Improved |
| use distribution-based sampling (KDE, GAN) instead of neighborhood sampling (ablation) | sampling · MNIST | test accuracy (%) | Improved |
| transfer SVHN to MNIST with KDE30 samples vs scratch training, epoch 1 | S_KDE30 · SVHN to MNIST | test accuracy (%): 20.6 -> 86.9 | Improved |
| S_KDE30 vs pre-train and fine-tune B_F, epoch 1 | S_KDE30 · SVHN to MNIST | test accuracy (%): 84.4 -> 86.9 | Improved |
Closest papers by meaning
- Towards Scalable and Versatile Weight Space Learning (ICML 2024 · similarity 0.89)
- HyperNet Fields: Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories (CVPR 2025 · similarity 0.87)
- Foundation Models Secretly Understand Neural Network Weights: Enhancing Hypernetwork Architectures with Foundation Models (ICLR 2025 · similarity 0.87)
- Hyper-Transforming Latent Diffusion Models (ICML 2025 · similarity 0.86)
- From Lazy to Rich: Exact Learning Dynamics in Deep Linear Networks (ICLR 2025 · similarity 0.85)
- Tracing Representation Progression: Analyzing and Enhancing Layer-Wise Similarity (ICLR 2025 · similarity 0.84)
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.
