-
摘要:
当前,市场上普遍使用的负责推理的终端人工智能(AI)芯片使用训练好的参数对数据进行快速高效运算。但在通常训练过程中使用的数据集和真实数据的分布不一致,由此获得的参数会导致终端AI芯片识别准确度降低。为此,提出了一种基于终端AI芯片的可视化反馈系统架构方法。使用反卷积特征可视化方法,在具有高效计算性能的终端AI芯片上,对卷积核参数进行迭代优化,达到可识别该图像目的。相比于CPU/GPU和FPGA,所提架构在卷积神经网络模型里,更具有高效处理能力和灵活可塑性。实验表明,该研究有效提高了终端AI芯片的普适性、识别准确度和处理效率。
-
关键词:
- 深度学习 /
- 终端人工智能(AI)芯片 /
- 卷积层可视化 /
- 卷积核参数优化 /
- 小样本
Abstract:Currently, terminal Artificial Intelligence (AI) chips responsible for inference are commonly used in the market, which use trained parameters to perform fast and efficient calculations on data. However, the training dataset usually has different distribution with the real-world data, and the parameters obtained in this case lead to a decrease in the accuracy of the terminal chip recognition. To this end, this paper proposes a visual feedback system architecture method based on terminal AI chip. Using the deconvolution feature visualization method, the convolution kernel parameters are iteratively optimized on the terminal AI chip with high computational performance to achieve the purpose of recognizing the image. Compared with CPU/GPU and FPGA, the architecture proposed in this paper has more efficient processing capability and flexible plasticity in the convolutional neural network model. Experiments show that the research effectively improves the universality, recognition accuracy and handling efficiency of the terminal chip.
-
表 1 不同模型实验结果对比
Table 1. Comparison of experimental results among different models
模型 优化前识别率/% 优化后识别率/% 平均处理时间/ms 平均优化参数时间/s CPU GPU FPGA 本文架构 CPU GPU FPGA 本文架构 AlexNet 54.71 56.26 1 315.1 26.3 3.87 1.94 75 938.3 1 518.65 371.8 48.88 ResNet18 65.39 67.12 5 720.2 161.8 12.84 5.26 188 326.9 22 003.5 725.3 122.29 ResNet50 66.82 67.87 21 736.95 441.8 38.15 14.08 327 688.8 54 300.6 1 964.5 301.73 ResNet101 67.63 68.03 39 126.51 994.1 67.03 25.6 701 909.4 101 542.1 4 553.7 638.59 YOLO3 54.95 56.42 169 065.6 4 682.5 289.64 116.28 2 816 060.9 459 684.2 22 003.5 2 701.6 -
[1] HINTON G E, SALAKHUTDINOV R.Reducing the dimensionality of data with neural networks[J].Science, 2006, 313(5786):504-507. doi: 10.1126/science.1127647 [2] MIROWSKI P, LECUN Y, MADHAVAN D, et al.Comparing SVM and convolutional networks for epileptic seizure prediction from intracranial EEG[C]//Proceedings of IEEE International Workshop on Machine Learning for Signal Processing.Piscataway: IEEE Press, 2008: 244-249. [3] 杨喜权.数字逻辑电路[M].成都:电子科技大学出版社, 2004:150.YANG X Q.Digital logic circuit[M].Chengdu:University of Electronic Science and Technology Press, 2004:150(in Chinese) [4] 施羽暇.人工智能芯片技术研究[J].电信网技术, 2016(12):11-13. http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=dxwjs201612003SHI Y X.Research on artificial intelligence process chip technology[J].Telecommunications Network Technology, 2016(12):11-13(in Chinese). http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=dxwjs201612003 [5] HECHT-NIELSEN R.Theory of the backpropagation neural network[J].Neural Networks, 1988, 1(S1):445-448. http://cn.bing.com/academic/profile?id=3d38f7b0ec99dd230a147079cb5d6bd0&encoded=0&v=paper_preview&mkt=zh-cn [6] 张美芳, 王羽, 郑碧琪, 等.人工智能在汽车自动驾驶中的应用[J].汽车工业研究, 2019(3):2-7. doi: 10.3969/j.issn.1009-847X.2019.03.001ZHANG M F, WANG Y, ZHENG B Q, et al.Application of artificial intelligence in autonomous driving[J].Automotive Industry Research, 2019(3):2-7(in Chinese). doi: 10.3969/j.issn.1009-847X.2019.03.001 [7] 张顺, 龚怡宏, 王进军.深度卷积神经网络的发展及其在计算机视觉领域的应用[J].计算机学报, 2019, 42(3):453-482. http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=jsjxb201903001ZHANG S, GONG Y H, WANG J J.The development of deep convolutional neural network and its application in the field of computer vision[J].Chinese Journal of Computers, 2019, 42(3):453-482(in Chinese). http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=jsjxb201903001 [8] BURDEA G C, COIFFET P.Virtual reality technology[M].New York:John Wiley & Sons, 2003:12-13. [9] KRIZHEVSKY A, SUTSKEVER I, HINTON G E, et al.ImageNet classification with deep convolutional neural networks[J].Neural Information Processing Systems, 2012, 141(5):1097-1105. http://cn.bing.com/academic/profile?id=1dc5d01904d2c274eaec2181a93aa339&encoded=0&v=paper_preview&mkt=zh-cn [10] KOTHARI S C, OH H.Neural networks for pattern recognition[J].Advances in Computers, 1993, 37:119-166. doi: 10.1016/S0065-2458(08)60404-0 [11] KRIZHEVSKY A, ILYA S, HINTON G E.Imagenet classification with deep convolutional neural networks[C]//Proceedings of the 25th International Conference on Neural Information Processing Systems-Volume 1.Red Hook: Curran Associates Inc., 2012: 1097-1105. [12] SIMONYAN K, ZISSERMAN A.Very deep convolutional networks for large-scale image recognition[EB/OL].(2015-04-10)[2019-09-20].https://arxiv.org/abs/1409.1556. [13] SZEGEDY C, LIU W, JIA Y, et al.Going deeper with convolutions[C]//IEEE Conference on Computer Vision and Pattern Recognition.Piscataway: IEEE Press, 2015: 1-9. [14] HE K, ZHANG X, REN S, et al.Deep residual learning for image recognition[C]//IEEE Conference on Computer Vision and Pattern Recognition.Piscataway: IEEE Press, 2016: 770-778. [15] HU J, SHEN L, SUN G, et al.Squeeze-and-excitation networks[EB/OL].(2019-05-16)[2019-09-20].https://arxiv.org/abs/1709.01507. [16] REDMON J, FARHADI A.YOLOv3: An incremental improvement[EB/OL].(2018-04-08)[2019-09-20].https://arxiv.org/abs/1804.02767. [17] ZEILER M D, FERGUS R.Visualizing and understanding convolutional networks[C]//European Conference on Computer Vision.Berlin: Springer, 2014: 818-833. [18] ZEILER M D, TAYLOR G W, FERGUS R, et al.Adaptive deconvolutional networks for mid and high level feature learning[C]//International Conference on Computer Vision.Piscataway: IEEE Press, 2011: 2018-2025. [19] GOODFELLOW I, POUGET-ABADIE J, MIRZA M, et al.Generative adversarial nets[C]//Advances in Neural Information Processing Systems.Red Hook: Curran Associates Inc., 2014: 2672-2680.