Research map / Architectures and efficiency / Quantization
Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations
Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, Yoshua Bengio
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 |
|---|---|---|---|
| widen the hard tanh boundaries after training and retrain (ablation) | BNN on AlexNet · ImageNet | top-1 accuracy: improve top-1 accuracies by 1.5% | Improved |
| use 2-bit activations instead of 1-bit with binary weights | BNN on AlexNet · ImageNet | top-1 accuracy: 41.8% -> 51.03% | Improved |
| use 4-bit weights and activations (QNN) instead of binary | GoogleNet · ImageNet | top-1 accuracy: 47.1% -> 66.5% | Improved |
| also quantize gradients to 6 bits (QNN 6-bit) | GoogleNet QNN · ImageNet | top-1 accuracy: 66.5% -> 66.4% | No clear effect |
| replace vanilla BN and Adam with shift-based versions | BNN · MNIST/CIFAR-10 | accuracy | No clear effect |
| binarize RNN weights and activations | RNN/LSTM on Penn Treebank · Penn Treebank | accuracy | Hurt |
| increase RNN to 4-bit weights and activations | LSTM 300 units on Penn Treebank · Penn Treebank | perplexity per word: 220 -> 100 | Improved |
| use XNOR GPU kernel instead of baseline kernel | binary MLP on MNIST · MNIST | runtime: 7 times faster | Improved |
| use XNOR kernel for 8192x8192x8192 binary matmul | GPU kernel · GTX750 matmul | speedup: 23 times faster than baseline, 3.4 times faster than cuBLAS | Improved |
Closest papers by meaning
- Training Quantized Nets: A Deeper Understanding (NeurIPS 2017 · similarity 0.91)
- Scalable methods for 8-bit training of neural networks (NeurIPS 2018 · similarity 0.90)
- Linear Symmetric Quantization of Neural Networks for Low-precision Integer Hardware (ICLR 2020 · similarity 0.90)
- LQ-Nets: Learned Quantization for Highly Accurate and Compact Deep Neural Networks (ECCV 2018 · similarity 0.90)
- Analysis of Quantized Models (ICLR 2019 · similarity 0.90)
- Searching for Low-Bit Weights in Quantized Neural Networks (NeurIPS 2020 · similarity 0.90)
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.
