Research map / Label-efficient and robust learning / Recognition and long-tail
Lossy Compression for Lossless Prediction
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 CLIP+EB instead of JPEG | zero-shot image classification compressor · ImageNet | rate gain vs JPEG: $1104\times$ | Helped |
| Use CLIP+EB instead of JPEG | zero-shot image classification compressor · 8 zero-shot datasets | classification accuracy | No clear effect |
| Add an entropy bottleneck to CLIP representations (ablation) | CLIP compressor · ImageNet | bit-rate per image: 1.52e4 -> 1.35e3 | Helped |
| Increase entropy-bottleneck β from the standard setting to high β (ablation) | CLIP compressor · ImageNet | test accuracy (%): 76.5 -> 76.6 | No clear effect |
| Use low β for the entropy bottleneck (ablation) | CLIP compressor · ImageNet | test accuracy (%): 76.5 -> 76.0 | Hurt |
| Train BINCE end-to-end rather than staggered (ablation) | BINCE on augmented MNIST · augmented MNIST | bit-rate: 358 to 131 bits | Helped |
| Predict from representations rather than reconstructions (ablation) | VIC · VIC downstream prediction | accuracy: +9% | Helped |
| Predict from representations rather than reconstructions (ablation) | VC · VC downstream prediction | accuracy: -12% | Hurt |
| Use text-image invariance (CLIP) instead of standard image augmentations (SimCLR) (ablation) | SSL compressor · ImageNet | bit-rate per image: $2811$ -> $2108$ | Helped |
| Use VIC rather than standard neural compression | neural image compressor · Banana source | AURD: $48.1 { \pm 0.3}$ -> $35.8 { \pm 4.2}$ | Helped |
| Use stronger information-discarding augmentations while preserving labels (ablation) | VIC on MNIST · MNIST | rate: 185.3, 79.0, and 5.7 bits | Helped |
| Use BINCE instead of PNG lossless compression | STL10 classifier compressor · STL10 | bit-rate gain: 121× | Helped |
Closest papers by meaning
- Real-Time Adaptive Image Compression (ICML 2017 · similarity 0.88)
- Full Resolution Image Compression With Recurrent Neural Networks (CVPR 2017 · similarity 0.88)
- Learning Better Lossless Compression Using Lossy Compression (CVPR 2020 · similarity 0.88)
- Practical Full Resolution Learned Lossless Image Compression (CVPR 2019 · similarity 0.88)
- Optimal Neural Compressors for the Rate-Distortion-Perception Tradeoff (NeurIPS 2025 · similarity 0.87)
- Towards Backward-Compatible Continual Learning of Image Compression (CVPR 2024 · similarity 0.87)
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.