Research map / Learning theory and algorithms / Prediction and decision losses
Deconstructing Data Reconstruction: Multiclass, Weight Decay and General Losses
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Train on multiple classes rather than binary labels | MLP D-1000-1000-C on CIFAR10 · CIFAR10, 500 training samples | good reconstructions (SSIM>0.4): 40 -> 101 | Helped |
| Increase number of classes C in {2,3,4,5,10} at fixed samples per class | MLP classifiers · CIFAR10 | good reconstructions | Helped |
| Add weight decay during training with standard initialization | MLP classifiers, C=2 and 10, 50 samples per class · CIFAR10 | good reconstructions | Helped |
| Add weight decay to a CNN (Conv k=3, 32 channels) | Conv-1000-1 binary classifier · CIFAR10 | good reconstructions | Helped |
| Train with regression losses (MSE, L2.5, Huber) instead of classification losses | Reconstruction scheme · CIFAR10 binary | SSIM | No clear effect |
| Increase training set size N to 500 even with larger width W | MLP D-W-W-1 · CIFAR10 binary | fraction of samples reconstructed: 8 | Hurt |
| Scale to W=10,000 and N=5,000 | MLP D-W-W-1 · CIFAR10 binary | reconstruction quality | Mixed |
Closest papers by meaning
- Reconstructing Training Data From Trained Neural Networks (NeurIPS 2022 · similarity 0.90)
- A Law of Data Reconstruction for Random Features (And Beyond) (ICLR 2026 · similarity 0.86)
- Wide Neural Networks Trained with Weight Decay Provably Exhibit Neural Collapse (ICLR 2025 · similarity 0.86)
- Understanding Reconstruction Attacks with the Neural Tangent Kernel and Dataset Distillation (ICLR 2024 · similarity 0.85)
- Thumb on the Scale: Optimal Loss Weighting in Last Layer Retraining (NeurIPS 2025 · similarity 0.85)
- Understanding and Leveraging the Learning Phases of Neural Networks (AAAI 2024 · similarity 0.85)
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.