文章详情

题名 · 作者 · 摘要 · 关键词 · DOI · 参考文献

深度神经网络低比特量化方法综述

田程，李正杰，陈功富，赵明，冯树萱

(成都华微电子科技股份有限公司，四川 成都 610041)

栏目：人工智能 · 第 2025 年第 10 期 · 页码：23-33,38 · 2096-4706（2025）10-0023-12 · 中图分类号：TP391 · 文献标识码：A · DOI：10.19850/j.cnki.2096-4706.2025.10.006

摘要

近年来，深度神经网络(Deep Neural Networks, DNNs)在计算机视觉和自然语言处理等领域都取得了突破性进展。然而，DNNs通常包含大量参数，并消耗大量的计算资源，这限制了它们在资源受限设备上的部署与应用。为了减少内存占用和计算负担，有关模型压缩与加速的研究工作不断涌现，模型低比特量化是其中主要的方法之一。模型低比特量化旨在将原始高精度的FP32的浮点运算替代为效率更高的低精度(≤8 bit)定点或位运算，从而大大减少模型的计算成本，使量化后的网络能够部署在边缘设备上。针对计算机视觉领域的图像分类与目标检测任务，全面调研模型低比特量化技术的相关现状，并深入分析这些方法的优缺点；此外，将概述深度神经网络低比特量化基础，并对比部分代表性的方法在不同数据集上的性能表现；最后，对模型低比特量化的未来发展趋势进行展望。

关键词

深度神经网络；模型压缩与加速；低比特量化

参考文献

  1. [1] KRIZHEVSKY A，SUTSKEVER I，HINTON G E. Imagenet Classification with Deep Convolutional Neural Networks [J].Advances in Neural Information Processing Systems，2017，60（6）：84-90.
  2. [2] ADARSH P，RATHI P，KUMAR M. YOLO v3-Tiny: Object Detection and Recognition Using One Stage Improved Model [C]//2020 6th International Conference on Advanced Computing and Communication Systems(ICACCS).Coimbatore：IEEE，2020：687-694.
  3. [3] BERTON G，MEREU R，TRIVIGNO G，et al. Deep Visual Geo-Localization Benchmark [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.New Orleans：IEEE，2022：5396-5407.
  4. [4] SAITO K，SOHN K，ZHANG X，et al. Pic2word: Mapping Pictures to Words for Zero-Shot Composed Image Retrieval [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Vancouver：IEEE，2023：19305-19314.
  5. [5] QI D，TAO T，ZHANG Z，et al. AMVP: Adaptive Multi-Volume Primitives for Auto-Driving Novel View Synthesis [J].IEEE Robotics and Automation Letters，2024，9（10）：8306-8313.
  6. [6] SPAGNOLO F，CORSANELLO P，FRUSTACI F，et al. Design of a Low-Power Super-Resolution Architecture for Virtual Reality Wearable Devices [J].IEEE Sensors Journal，2023，23（8）：9009-9016.
  7. [7] KONG Q R，WU Y R，YUAN C，et al. CT-CAD: Context-Aware Transformers for End-to-End Chest Abnormality Detection on X-Rays [C]//IEEE/ACM Transactions on Computational Biology and Bioinformatics.Houston：2023， IEEE，2023：1385-1388.
  8. [8] SIMONYAN K. Very Deep Convolutional Networks for Large-Scale Image Recognition [J/OL].arXiv:1409.1556 [cs.CV].（2014-09-04）.[2024-10-19].https://arxiv.org/abs/1409.1556.
  9. [9] ZHOU A J，MA Y K，ZHU J N，et al. Learning N: M Fine-grained Structured Sparse Neural Networks From Scratch [J/OL].arXiv:2102.04010 [cs.CV].（2021-04-18）.[2024-10-19].https://arxiv.org/abs/2102.04010.
  10. [10] PHAM C，HOANG T，DO T T. Collaborative Multi-Teacher Knowledge Distillation for Learning Low Bit-Width Deep Neural Networks [C]//Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. Waikoloa：IEEE，2023：6435-6443.
  11. [11] SARAGADAM V，BALESTRIERO R，VEERARAGHAVAN A，et al. DeepTensor:Low-Rank Tensor Decomposition with Deep Network Priors [J].IEEE Transactions on Pattern Analysis and Machine Intelligence，2024，46（12）：10337-10348.
  12. [12] XUE Y，CHEN C，SLOWIK A. Neural Architecture Search Based on a Multi-Objective Evolutionary Algorithm with Probability Stack [J].IEEE Transactions on Evolutionary Computation，2023，27（4）：778-786.
  13. [13] JACOB B，KLIGYS S，CHEN B，et al. Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.Salt Lake City：IEEE，2018：2704-2713.
  14. [14] LIAO Z，QUÉTU V，NGUYEN V T，et al. Can Unstructured Pruning Reduce the Depth in Deep Neural Networks? [C]//Proceedings of the IEEE/CVF International Conference on Computer Vision.Paris：IEEE，2023：1402-1406.
  15. [15] REED J K，DEVITO Z，HE H，et al. Torch.fx: Practical Program Capture and Transformation for Deep Learning in Python [J/OL].arXiv:2112.08429 [cs.LG].（2021-12-15） [2022-12-27].https://arxiv.org/abs/2112.08429.
  16. [16] DENG J，DONG W，SOCHER R，et al. Imagenet: A Large-Scale Hierarchical Image Database [C]//2009 IEEE Conference on Computer Vision and Pattern Recognition.Miami：IEEE，2009：248-255.
  17. [17] ZHANG D，YANG J，YE D，et al. LQ-Nets: Learned Quantization for Highly Accurate and Compact Deep Neural Networks [C]//Proceedings of the European Conference on Computer Vision (ECCV).Munich：Springer，2018：365-382.
  18. [18] ZHOU A，YAO A，GUO Y，et al. Incremental Network Quantization：Towards Lossless Cnns with Low-Precision Weights [J/OL].arXiv:1702.03044 [cs.CV].（2017-02-10）[2024-10-19].https://arxiv.org/abs/1702.03044.
  19. [19] ZHOU S，WU Y，NI Z，et al. Dorefa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients [J/OL].arXiv:1606.06160 [cs.NE].（2016-06-20）[2024-10-19].https://arxiv.org/abs/1606.06160.
  20. [20] LI F，LIU B，WANG X，et al. Ternary Weight Networks [C]//ICASSP 2023-2023 IEEE International Conference on Acoustics，Speech and Signal Processing (ICASSP).Rhodes Island：IEEE，2023：1-5.
  21. [21] GONG Y，LIU L，YANG M，et al. Compressing Deep Convolutional Networks Using Vector Quantization [J/OL].arXiv:1412.6115 [cs.CV].（2014-12-10）[2024-10-19].https://arxiv.org/abs/1412.6115.
  22. [22] STOCK P，JOULIN A，GRIBONVAL R，et al. And the Bit Goes Down: Revisiting the Quantization of Neural Networks [J/OL].arXiv:1907.05686 [cs.CV].（2020-11-09）[2024-10-19].https://arxiv.org/abs/1907.05686.
  23. [23] LI Y，DONG X，WANG W. Additive Powers-of-Two Quantization：An Efficient Non-Uniform Discretization for Neural Networks [J/OL].arXiv:1909.13144 [cs.LG].（2019-08-28）[2024-10-19].https://arxiv.org/abs/1909.13144.
  24. [24] FARAO J，FRASER N，BLOTT M，et al. SYQ: Learning Symmetric Quantization for Efficient Deep Neural Networks [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.Salt Lake City：IEEE，2018：4300-4309.
  25. [25] SZEGEDY C，LIU W，JIA Y，et al. Going Deeper with Convolutions [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.Boston：IEEE，2015：1-9.
  26. [26] HE K，ZHANG X，REN S，et al. Deep Residual Learning for Image Recognition[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.Las Vegas，IEEE：2016：770-778.
  27. [27] IANDOLA F N. SqueezeNet: AlexNet-Level Accuracy with 50x Fewer Parameters and< 0.5 MB Model Size [J/OL].arXiv:1602.07360 [cs.CV].（2016-11-04）[2024-10-19].https://arxiv.org/abs/1602.07360.
  28. [28] LECUN Y，BOTTOU L，BENGIO Y，et al. Gradient-Based Learning Applied to Document Recognition [J].Proceedings of the IEEE，1998，86（11）：2278-2324.
  29. [29] CHEN W，WILSON J，TYREE S，et al. Compressing Neural Networks with the Hashing Trick [C]//International Conference on Machine Learning.Lille：PMLR，2015：2285-2294.
  30. [30] ZHANG A，YANG Z，WANG N，et al. Comq:A Backpropagation-Free Algorithm for Post-Training Quantization [J/OL].arXiv:2403.07134 [cs.LG].（2024-03-11）[2024-10-19].https://arxiv.org/abs/2403.07134.
  31. [31] LI X，LIU B，YANG R H，et al. DenseShift: Towards Accurate and Efficient Low-Bit Power-of-Two Quantization [C]//Proceedings of the IEEE/CVF International Conference on Computer Vision.Las Vegas：IEEE，2023：17010-17020.
  32. [32] LI J，CHEN G，JIN M，et al. AE-Qdrop: Towards Accurate and Efficient Low-Bit Post-Training Quantization for a Convolutional Neural Network [J].Electronics，2024，13（3）：644.
  33. [33] SANDLER M，HOWARD A，ZHU M，et al. Mobilenetv2: Inverted Residuals and Linear Bottlenecks [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City：IEEE，2018：4510-4520.
  34. [34] TAN M，LE Q. Efficientnet: Rethinking Model Scaling for Convolutional Neural Networks [C]//International Conference on Machine Learning.PMLR，2019：6105-6114.
  35. [35] BANNER R，NAHSHAN Y，SOUDRY D. Post Training 4-Bit Quantization of Convolutional Networks for Rapid-Deployment [J/OL].arXiv:1810.05723 [cs.CV].（2018-10-02）[2024-10-19].https://arxiv.org/abs/1810.05723.
  36. [36] LOU Q，GUO F，LIU L，et al. AutoQ: Automated Kernel-Wise Neural Network Quantization [J].arXiv:1902.05690 [cs.LG].（2019-02-15）[2024-10-19].https://arxiv.org/abs/1902.05690.
  37. [37] XU K，LI Z，WANG S，et al. PTMQ: Post-Training Multi-Bit Quantization of Neural Networks [C]//Proceedings of the AAAI Conference on Artificial Intelligence.Washington：AAAI Press，2024：16193-16201.
  38. [38] YANG W，ZHI X，TONG W. Optimization of Linear Quantization for General and Effective Low Bit-Width Network Compression [J/OL].Algorithms，2023，16（1）：31（2023-01-04）[2024-10-19].https://doi.org/10.3390/a16010031.
  39. [39] CHAUHAN A，TIWARI U. Post Training Mixed Precision Quantization of Neural Networks Using First-Order Information [C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. Paris：IEEE，2023：1343-1352.
  40. [40] IANDOLA F N，HAN S，MOSKEWICZ M W，et al. SqueezeNet: Alexnet-Level Accuracy with 50x Fewer Parameters and<0.5 MB Model Size [J/OL].arXiv:1602.07360 [cs.CV].（2016-02-24）[2024-10-19].https://arxiv.org/abs/1602.07360.
  41. [41] SZEGEDY C，VANHOUCKE V，IOFFE S，et al. Rethinking the Inception Architecture for Computer Vision [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.Las Vegas：IEEE，2016：2818-2826.
  42. [42] XU K，HAN L，TIAN Y，et al. Eq-net: Elastic Quantization Neural Networks [C]//Proceedings of the IEEE/CVF International Conference on Computer Vision.Paris：IEEE，2023：1505-1514.
  43. [43] BENGIO Y，LÉONARD N，COURVILLE A. Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation [J/OL].arXiv:1308.3432 [cs.LG].（2013-08-15）[2024-10-19].https://arxiv.org/abs/1308.3432.
  44. [44] XU Q，CHEN Z，RAGAB M，et al. Contrastive Adversarial Knowledge Distillation for Deep Model Compression in Time-Series Regression Tasks [J].Neurocomputing，2022，485：242-251.
  45. [45] QU Z，ZHOU Z，CHENG Y，et al. Adaptive Loss-Aware Quantization for Multi-Bit Networks [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Seattle：IEEE，2020：7988-7997.
  46. [46] YANG L，JIN Q. Fracbits: Mixed Precision Quantization via Fractional Bit-Widths [C]//Proceedings of the AAAI Conference on Artificial Intelligence.AAAI Press，2021，35（12）：10612-10620.
  47. [47] ZHAO S，YUE T，HU X. Distribution-Aware Adaptive Multi-Bit Quantization [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Nashville：IEEE，2021：9281-9290.
  48. [48] SHER A，TRUSOV A，LIMONOVA E，et al. Neuron-by-Neuron Quantization for Efficient Low-Bit QNN Training [J].Mathematics，2023，11（9）：2112.
  49. [49] LUO R，DRIMBAREAN A，MCDERMOTT J，et al. Reclaiming Residual Knowledge: A Novel Paradigm to Low-Bit Quantization [J/OL].arXiv:2408.00923 [cs.CV].（2024-08-01）[2024-10-19].https://arxiv.org/abs/2408.00923.
  50. [50] ELTHAKEB A T，PILLIGUNDLA P，MIRESHGHALLAGH F S，et al. ReLeQ: A Reinforcement Learning Approach for Deep Quantization of Neural Networks [J].IEEE Micro，2020，40（5）：37-45.
  51. [51] WANG K，LIU Z，LIN Y，et al. HAQ: Hardware-Aware Automated Quantization with Mixed Precision [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Long Beach：IEEE，2019：8612-8620.
  52. [52] WU B，WANG Y，ZHANG P，et al. Mixed Precision Quantization of Convnets via Differentiable Neural Architecture Search [J/OL].arXiv:1812.00090 [cs.CV].（2018-11-30）[2024-10-19].https://arxiv.org/abs/1812.00090.
  53. [53] DONG Z，YAO Z，GHOLAMI A，et al. HAWQ: Hessian Aware Quantization of Neural Networks with Mixed-Precision [C]//Proceedings of the IEEE/CVF International Conference on Computer Vision.Seoul：IEEE，2019：293-302.
  54. [54] LIU W，ANGUELOV D，ERHAN D，et al. SSD: Single Shot Multibox Detector [C]//Computer Vision–ECCV 2016.Amsterdam:Springer，2016：21-37.
  55. [55] REDMON J. You Only Look Once: Unified, Real-Time Object Detection [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas：IEEE，2016：779-788.
  56. [56] HE K，GKOIXARI G，DOLLÁR P，et al. Mask R-CNN [C]//Proceedings of the IEEE International Conference on Computer Vision.Venice：IEEE，2017：2961-2969.
  57. [57] LIN T Y，MAIRE M，BELONGIE S，et al. Microsoft COCO: Common Objects in Context[C]//Computer Vision-ECCV 2014.Zurich：Springer International Publishing，2014：740-755.
  58. [58] GUPTA K，ASTHANA A. Reducing the Side-Effects of Oscillations in Training of Quantized YOLO Networks [C]//Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision.Waikoloa：IEEE，2024：2452-2461.
  59. [59] ESSER S K，MCKINSTRY J L，BABLANI D，et al. Learned Step Size Quantization [J/OL].arXiv:1902.08153 [cs.LG].（2019-02-21）[2024-10-19].https://arxiv.org/abs/1902.08153.
  60. [60] LI R，WANG Y，LIANG F，et al. Fully Quantized Network for Object Detection [C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).Long Beach：IEEE，2019：2805-2814.
  61. [61] CHEN P，LIU J，ZHUANG B，et al. AQD: Towards Accurate Quantized Object Detection [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Nashville：IEEE，2021：104-113.
  62. [62] ZHUANG B，LIU L，TAN M，et al. Training Quantized Neural Networks with a Full-Precision Auxiliary Module [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle：IEEE，2020：1488-1497.
  63. [63] BUCHNEV V，HE J，SUN F，et al. RUPQ: Improving Low-Bit Quantization by Equalizing Relative Updates of Quantization Parameters [C]//The 34th British Machine Vision Conference.Aberdeen：BMVA，2023：307-309.
  64. [64] ULTRALYTICS J. YOLOv5: A State-of-the-Art Real-Time Object Detection System [J/OL].[2024-10-19].https://iq.opengenus.org/yolov5/.
  65. [65] WANG C Y，BOCHKOVSKIY A，LIAO H Y M. YOLOv7: Trainable Bag-of-Freebies Sets New State-of-the-Art for Real-Time Object Detectors [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Vancouver：IEEE，2023：7464-7475.
  66. [66] LIN T Y，GOYAL P，GIRSHICK R，et al. Focal Loss for Dense Object Detection [J].IEEE Transactions on Pattern Analysis and Machine Intelligence，2020，42（2）：318-327.
  67. [67] SHI X，DING J，HAO Z，et al. Towards Energy Efficient Spiking Neural Networks: An Unstructured Pruning Framework [C]//The Twelfth International Conference on Learning Representations.Vienna：ICLR，2024：1-21.
  68. [68] GUO H，GREENGARD P，XING E P，et al. LQ-LoRa: Low-rank Plus Quantized Matrix Decomposition for Efficient Language Model Finetuning [J/OL].arXiv:2311.12023 [cs.CL].（2023-11-20）[2024-10-19].https://arxiv.org/abs/2311.12023.
  69. [69] ELSHAWI R，MAHER M，SAKR S. Automated Machine Learning: State-of-the-Art and Open Challenges [J/OL].arXiv:1906.02287 [cs.LG].（2019-06-05）[2024-10-19].https://arxiv.org/abs/1906.02287.

作者简介

田程(1998—)，女，汉族，贵州毕节人，视觉算法FPGA加速工程师，硕士，研究方向：模型量化与加速；李正杰(1983—)，男，汉族，四川成都人，高级工程师，博士，研究方向：视觉算法FPGA加速；陈功富(1996—)，男，汉族，河南信阳人，工程师，学士，研究方向：NPU设计；赵明(1987—)，男，汉族，四川射洪人，工程师，硕士，研究方向：AI算法的FPGA加速；冯树萱(2001—)，男，汉族，海南海口人，硕士在读，研究方向：AI芯片设计。