Research map / Architectures and efficiency / Pruning and efficient tuning
Weightless: Lossy weight encoding for deep neural network compression
Brandon Reagan, Udit Gupta, Bob Adolf, Michael Mitzenmacher, Alexander Rush, Gu-Yeon Wei, David Brooks
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 |
|---|---|---|---|
| Replace CSR/Huffman encoding with Bloomier-filter encoding and arithmetic coding | Pruned and clustered DNN weights · LeNet5 FC-0, DNS (MNIST) | compression factor: 333× vs 224× | Improved |
| Replace CSR/Huffman encoding with Bloomier-filter encoding and arithmetic coding | Pruned and clustered DNN weights · VGG-16 FC-1 (ImageNet) | compression factor: 85.8× vs 88.4× | Hurt |
| Use dynamic network surgery pruning instead of magnitude pruning (ablation) | Weightless · LeNet5 FC-0 (MNIST) | compression factor: 74.2× -> 496× | Improved |
| Use dynamic network surgery pruning instead of magnitude pruning (ablation) | Weightless · LeNet-300-100 FC-0 (MNIST) | compression factor: 60.1× -> 174× | Improved |
| Increase sparsity through stronger pruning (ablation) | Weightless encoding · LeNet5 FC-0 (MNIST) | compression factor | Improved |
| Retrain subsequent layers around deterministic Bloomier-filter false positives | Weightless · VGG-16 FC-0 (ImageNet) | top-1 accuracy: +2.0 percentage points | Improved |
| Use Weightless rather than lossless Deep Compression after simplification | Pruned and clustered models · LeNet-300-100, LeNet5, and VGG-16 | test error | No clear effect |
| Apply arithmetic coding to Bloomier-filter tables (ablation) | Weightless · VGG-16 FC-1 (ImageNet) | compression factor: 74.6× -> 85.8×; baseline 71.2× -> 88.4× | Hurt |
Later papers that compared with this method
1 later accepted paper reports Weightless as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Minimal Random Code Learning: Getting Bits Back from Compressed Model Parameters (ICLR 2019 · better on 2 of 2 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Lossy and Lossless (L2) Post-training Model Size Compression (ICCV 2023 · similarity 0.90)
- Weight Fixing Networks (ECCV 2022 · similarity 0.89)
- Loss-aware Weight Quantization of Deep Networks (ICLR 2018 · similarity 0.88)
- Minimal Random Code Learning: Getting Bits Back from Compressed Model Parameters (ICLR 2019 · similarity 0.88)
- Soft Weight-Sharing for Neural Network Compression (ICLR 2017 · similarity 0.87)
- On Compressing Deep Models by Low Rank and Sparse Decomposition (CVPR 2017 · 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.
