Research map / Learning theory and algorithms / Prediction and decision losses
Optimal Binary Autoencoding with Pairwise Correlations
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 |
|---|---|---|---|
| Use PC-AE with binary encodings instead of the vanilla single-layer AE with real encodings | MNIST · MNIST, H=32 and H=100 | cross-entropy reconstruction loss: 64.3 -> 51.9; 25.0 -> 9.2 | Helped |
| Use PC-AE with binary encodings instead of the vanilla single-layer AE with real encodings | Omniglot · Omniglot, H=32 and H=100 | cross-entropy reconstruction loss: 90.6 -> 76.1; 45.4 -> 12.1 | Helped |
| Use PC-AE with binary encodings instead of the vanilla single-layer AE with real encodings | Caltech-101 Silhouettes · Caltech-101 Silhouettes, H=32 and H=100 | cross-entropy reconstruction loss: 87.6 -> 54.5; 45.4 -> 7.1 | Helped |
| Use PC-AE with binary encodings instead of the vanilla single-layer AE with real encodings | notMNIST · notMNIST, H=32 and H=100 | cross-entropy reconstruction loss: 141.8 -> 121.9; 92.1 -> 62.2 | Mixed |
| Use binary rather than real-valued encodings (ablation) | PC-AE on MNIST · MNIST, H=32 and H=100 | cross-entropy reconstruction loss: 53.8 -> 51.9; 9.9 -> 9.2 | Helped |
| Use binary rather than real-valued encodings (ablation) | PC-AE on Omniglot · Omniglot, H=32 and H=100 | cross-entropy reconstruction loss: 77.2 -> 76.1; 13.2 -> 12.1 | Helped |
| Use binary rather than real-valued encodings (ablation) | PC-AE on Caltech-101 Silhouettes · Caltech-101 Silhouettes, H=32 and H=100 | cross-entropy reconstruction loss: 54.9 -> 54.5; 7.1 -> 7.1 | No clear effect |
| Use binary rather than real-valued encodings (ablation) | PC-AE on notMNIST · notMNIST, H=32 and H=100 | cross-entropy reconstruction loss: 122.4 -> 121.9; 63.0 -> 62.2 | Helped |
| Train the decoder on AE encodings using the pairwise-correlation decoder instead of the AE's learned decoder (ablation) | MNIST · MNIST, 32 and 100 hidden units | reconstruction loss: ~171/133 | Hurt |
| Train the decoder on AE encodings using the pairwise-correlation decoder instead of the AE's learned decoder (ablation) | Omniglot · Omniglot, 32 and 100 hidden units | reconstruction loss: ~211/134 | Hurt |
| Increase the hidden representation size from H=32 to H=100 (ablation) | PC-AE · MNIST, Omniglot, Caltech-101 Silhouettes, and notMNIST | cross-entropy reconstruction loss: MNIST: 51.9 -> 9.2; Omniglot: 76.1 -> 12.1; Caltech-101: 54.5 -> 7.1; notMNIST: 121.9 -> 62.2 | Helped |
Closest papers by meaning
- Learning Linear Block Error Correction Codes (ICML 2024 · similarity 0.84)
- Compression of Structured Data with Autoencoders: Provable Benefit of Nonlinearities and Depth (ICML 2024 · similarity 0.83)
- Approaching Rate-Distortion Limits in Neural Compression with Lattice Transform Coding (ICLR 2025 · similarity 0.83)
- Optimal Neural Compressors for the Rate-Distortion-Perception Tradeoff (NeurIPS 2025 · similarity 0.82)
- A Global Geometric Analysis of Maximal Coding Rate Reduction (ICML 2024 · similarity 0.82)
- Towards Lossless Implicit Neural Representation via Bit Plane Decomposition (CVPR 2025 · similarity 0.82)
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.