一种基于群等变卷积的度量元学习算法

doi:10.19678/j.issn.1000-3428.0060046

摘要/Abstract

摘要： 传统机器学习方法泛化性能不佳，需要通过大规模数据训练才能得到较好的拟合结果，因此不能快速学习训练集外的少量数据，对新种类任务适应性较差，而元学习可实现拥有类似人类学习能力的强人工智能，能够快速适应新的数据集，弥补机器学习的不足。针对传统机器学习中的自适应问题，利用样本图片的局部旋转对称性和镜像对称性，提出一种基于群等变卷积神经网络（G-CNN）的度量元学习算法，以提高特征提取能力。利用G-CNN构建4层特征映射网络，根据样本图片中的局部对称信息，将支持集样本映射到合适的度量空间，并以每类样本在度量空间中的特征平均值作为原型点。同时，通过同样的映射网络将查询机映射到度量空间，根据查询集中样本到原型点的距离完成分类。在Omniglot和miniImageNet数据集上的实验结果表明，该算法相比孪生网络、关系网络、MAML等传统4层元学习算法，在平均识别准确率和模型复杂度方面均具有优势。

关键词: 元学习, 群等变卷积, 深度学习, 自适应性, 度量学习

Abstract: Traditional machine learning methods have poor generalization performance and need large-scale data training to get better fitting results.They cannot quickly learn a small amount of data outside the training set, and have poor adaptability to new types of tasks.Meta-learning can realize strong artificial intelligence systems with similar learning ability to human beings.These artificial intelligence systems can quickly adapt to new data sets and make up for the shortcomings of machine learning.In order to solve the adaptibility problem of traditional machine learning, a metric meta-learning algorithm based on Group equivariant Convolution Neural Network(G-CNN) is proposed by using the local rotation symmetry and mirror symmetry of sample images to improve feature extraction ability.The G-CNN is used to form a 4-layer feature mapping network.According to the local symmetry information in the sample picture, the support set samples are mapped to the appropriate metric space, and the average feature of each kind of samples in the metric space is used as the prototype point.At the same time, the query machine is mapped to the metric space through the same mapping network, so as to complete the classification according to the distance between the sample in the query set and the prototype point.Experimental results on Omniglot and miniImageNet data sets show that the proposed algorithm displays an obvious advantage in average recognition accuracy and model complexity compared with traditional 4-layer meta learning algorithms, such as Siamese network, relational network and MAML.

Key words: meta-learning, group equivariant convolution, deep learning, adaptability, metric learning

中图分类号:

TP18

吴鹏翔, 李凡长. 一种基于群等变卷积的度量元学习算法[J]. 计算机工程, 2022, 48(2): 72-78.

WU Pengxiang, LI Fanzhang. A Metric Meta-Learning Algorithm Based on Group Equivariant Convolution[J]. Computer Engineering, 2022, 48(2): 72-78.

https://www.ecice06.com/CN/Y2022/V48/I2/72

图/表 9

20220226124523

20220226124526

20220226124530

20220226124534

20220226124537

20220226124540

20220226124544

20220226124549

20220226124554

参考文献

[1] JIN C, YANG Z, WANG Z, et al.Provably efficient reinforcement learning with linear function approximation[C]//Proceedings of the 33rd Annual Conference on Learning Theory.Graz, Austria:[s.n.], 2020:2137-2143.
[2] SILVER D, SCHRITTWIESER J, SIMONYAN K, et al.Mastering the game of go without human knowledge[J].Nature, 2017, 550(7676):354-359.
[3] ERIA K, JAYABALAN M.Neural machine translation:a review of the approaches[J].Journal of Computational and Theoretical Nanoscience, 2019, 16(8):3596-3602.
[4] SETHI A, GU M, GUMUSGOZ E, et al.Supervised enhancer prediction with epigenetic pattern recognition and targeted validation[J].Nature Methods, 2020, 17(8):807-814.
[5] LAI N, KAN M, HAN C, et al.Learning to learn adaptive classifier-predictor for few-shot learning[J].IEEE Transactions on Neural Networks and Learning Systems, 2021, 32(8):3458-3470.
[6] MISHRA N, ROHANINEJAD M, CHEN X, et al.A simple neural attentive meta-learner[EB/OL].(2017-07-11)[2020-11-10].https://arxiv.org/pdf/1707.03141v3.pdf.
[7] LI F F, FERGUS R, PERONA P.One-shot learning of object categories[J].IEEE Transactions on Pattern Analysis and Machine Intelligence, 2006, 28(4):594-611.
[8] SANTORO A, BARTUNOV S, BOTVINICK M, et al.Meta-learning with memory-augmented neural networks[C]//Proceedings of International Conference on Machine Learning.New York, USA:IEEE Press, 2016:1842-1850.
[9] KAISER Ł, NACHUM O, ROY A, et al.Learning to remember rare events[EB/OL].(2017-03-09)[2020-11-10].https://arxiv.org/pdf/1703.03129.pdf.
[10] FINN C, ABBEEL P, LEVINE S.Model-agnostic meta-learning for fast adaptation of deep networks[EB/OL].(2017-03-09)[2020-11-10].https://arxiv.org/pdf/1703.03400v3.pdf
[11] CHEN Z, FU Y, WANG Y X, et al.Image deformation meta-networks for one-shot learning[C]//Proceedings of 2019 IEEE Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2019:8680-8689.
[12] KOCH G, ZEMEL R, SALAKHUTDINOV R.Siamese neural networks for one-shot image recognition[C]//Proceedings of the 32nd International Conference on Machine Learning.Lille, France:[s.n.], 2015:1-8.
[13] VINYALS O, BLUNDELL C, LILLICRAP T, et al.Matching networks for one shot learning[C]//Proceedings of NIPS'16.Berlin, Germany:Springer, 2016:3630-3638.
[14] SNELL J, SWERSKY K, ZEMEL R.Prototypical networks for few-shot learning[C]//Proceedings of NIPS'17.Berlin, Germany:Springer, 2017:4077-4087.
[15] SUNG F, YANG Y, ZHANG L, et al.Learning to compare:relation network for few-shot learning[C]//Proceedings of 2018 IEEE Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2018:1199-1208.
[16] LING Y, XIAN Z, ZHOU Z.Holographic shear viscosity in hyperscaling violating theories without translational invariance[J].Journal of High Energy Physics, 2016(11):7.
[17] JIANG Y, LAI X, WATANABE K, et al.Charge order and broken rotational symmetry in magic-angle twisted bilayer graphene[J].Nature, 2019, 573(7772):91-95.
[18] NAMOZOV A, CHO Y I.An improvement for medical image analysis using data enhancement techniques in deep learning[C]//Proceedings of 2018 International Conference on Information and Communication Technology Robotics.Washington D.C., USA:IEEE Press, 2018:1-3.
[19] DOMMASCHK M, ECHAVARREN J, LEIGH D A, et al.Dynamic control of chiral space Through local symmetry breaking in a rotaxane organocatalyst[J].Angewandte Chemie International Edition, 2019, 58(42):14955-14958.
[20] COHEN T, WELLING M.Group equivariant convolutional networks[C]//Proceedings of International Conference on Machine Learning.New York, USA:ACM Press, 2016:2990-2999.
[21] VIALATTE J C, GRIPON V, MERCIER G.Generalizing the convolution operator to extend CNNs to irregular domains[EB/OL].(2016-06-03)[2020-11-10].https://arxiv.org/pdf/1606.01166.pdf.
[22] ULLRICH K, MEEDS E, WELLING M.Soft weight-sharing for neural network compression[EB/OL].(2017-02-13)[2020-11-10].https://arxiv.org/pdf/1702.04008.pdf.
[23] LI J, LI B, XU J, et al.Fully connected network-based intra prediction for image coding[J].IEEE Transactions on Image Processing, 2018, 27(7):3236-3247.
[24] JIANG W, HUANG C, DENG X.A new probability transformation method based on a correlation coefficient of belief functions[J].International Journal of Intelligent Systems, 2019, 34(6):1337-1347.
[25] CLEMONS E K, DEWAN R M, KAUFFMAN R J, et al.Understanding the information-based transformation of strategy and society[J].Journal of Management Information Systems, 2017, 34(2):425-456.
[26] LIU M Y, HUANG X, MALLYA A, et al.Few-shot unsupervised image-to-image translation[C]//Proceedings of 2019 IEEE International Conference on Computer Vision.Washington D.C., USA:IEEE Press, 2019:10551-10560.
[27] ETIKAN I, BALA K.Sampling and sampling methods[J].Biometrics & Biostatistics International Journal, 2017, 5(6):1-5.
[28] DEY N, CHEN A, GHAFURIAN S.Group equivariant generative adversarial networks[EB/OL].(2020-05-04)[2020-11-10].https://arxiv.org/pdf/2005.01683v2.pdf.

选择文件类型/文献管理软件名称

选择包含的内容