Research map / Architectures and efficiency / RNNs and normalization
Structured Multi-Hashing for Model Compression
Check 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 structured matrix-product multi-hashing instead of non-structured 1X, 2X, or 10X hashing | EfficientNet B2 · ImageNet, target size 5.3M | accuracy: 0.762 -> 0.774; 0.765 -> 0.774; 0.770 -> 0.774 | Helped |
| Use structured matrix-product multi-hashing instead of non-structured 1X, 2X, or 10X hashing | EfficientNet B2 · ImageNet, target size 7.9M | samples per second: 3900 -> 6040; 2500 -> 6040; 760 -> 6040 | Helped |
| Add per-layer learnable scale (ablation) | SMASH EfficientNet models · ImageNet, compressed EfficientNet B0–B4 | accuracy: 1.2%; 0.3%; 0.8%; 1.3%; 0.4%; 1.0%; 1.4%; 0.8%; 0.4%; 0.6%; 0.7%; 0.9%; 0.4%; 0.5%; 0.2% | Helped |
| Hash convolution weights only rather than all weights (ablation) | SMASH EfficientNet B1–B5 compressed to B0 size · ImageNet | accuracy | Mixed |
| Compress ResNet32 to 25% of its original variables using SMASH | ResNet32 · CIFAR10 | accuracy: 75% | No clear effect |
| Compress ResNet32 to 10% of its original size using SMASH | ResNet32 · CIFAR10 | accuracy: above 90% | Helped |
| Compress EfficientNet B4 to 2M parameters with SMASH | EfficientNet B4 · ImageNet | accuracy: 76.3% -> 76.6% | Helped |
| Compress EfficientNet B5 to 3M parameters with SMASH | EfficientNet B5 · ImageNet | accuracy: 78.8% -> 78.3% | Mixed |
| Apply SMASH rather than width multiplier | ResNet50 and ResNet101 · ImageNet | model size at 70% accuracy: half the size | Helped |
| Apply SMASH rather than width-multiplier or width-multiplier-plus-resolution baselines | MobileNetV2 · ImageNet | accuracy–model size trade-off | Helped |
| Apply SMASH rather than width-multiplier-plus-resolution near the full model | MobileNetV3 · ImageNet | accuracy–model size trade-off | Hurt |
Closest papers by meaning
- Hardware-Aware Compression with Random Operation Access Specific Tile (ROAST) Hashing (ICML 2023 · similarity 0.87)
- "Lossless" Compression of Deep Neural Networks: A High-dimensional Neural Tangent Kernel Approach (NeurIPS 2022 · similarity 0.87)
- From Hashing to CNNs: Training Binary Weight Networks via Hashing (AAAI 2018 · similarity 0.87)
- LSH-MoE: Communication-efficient MoE Training via Locality-Sensitive Hashing (NeurIPS 2024 · similarity 0.85)
- Forget the Data and Fine-Tuning! Just Fold the Network to Compress (ICLR 2025 · similarity 0.85)
- Network Memory Footprint Compression Through Jointly Learnable Codebooks and Mappings (ICLR 2024 · 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.