基于变分贝叶斯的纤维丛元学习算法  

Fiber Bundle Meta-learning Algorithm Based on Variational Bayes

作　　者：刘洋 李凡长[1] LIU Yang;LI Fan-zhang(School of Computer Science and Technology,Soochow University,Suzhou,Jiangsu 215006,China)

机构地区：[1]苏州大学计算机科学与技术学院,江苏苏州215006

出　　处：《计算机科学》2022年第3期225-231,共7页Computer Science

基　　金：国家重点研发计划(2018YFA0701700,2018YFA0701701);国家自然科学基金(61902269)。

摘　　要：以神经网络为基础的深度学习在大量领域取得优异成果,但其难以处理相似或未经训练的任务。深度学习在对新任务的学习和适应过程中存在困难,且对训练样本规模要求很高,造成泛化性和扩展性不佳的问题。元学习是一种新的学习框架,旨在解决传统学习方法难以解决的快速学习和适应新任务的问题。针对图像分类的元学习问题,文中提出了一种基于贝叶斯理论的纤维丛元学习算法(Fiber Bundle Meta-learning Algorithm,FBBML)。首先通过卷积神经网络提取支持数据集的图片信息,以得到图片的表示。然后构建数据特征的流形结构和数据特征到标签的纤维丛。最后输入查询集选取当前新任务的流形截面,从而获得适合新任务的纤维,得到图片的正确标签。实验结果表明,基于所提算法实现的模型(FBBML)在公共数据集(mini-ImageNet)上相比标准四层卷积神经网络的模型取得了最佳的准确率性能。同时将纤维丛理论引入元学习,使得算法本身具备更高的可解释性。Deep learning based on neural network has achieved excellent results in a large number of fields,but it is difficult to deal with similar or untrained tasks,and it is difficult to learn and adapt to new tasks.Moreover,it requires a high scale of trai-ning samples,resulting in its poor generalization and expansion.Meta learning is a new learning framework,which aims to solve the problem that traditional learning methods can’t solve fast learning and adapt to new tasks.Aiming at the meta learning problem of image classification,a novel fiber bundle meta learning algorithm based on Bayesian theory is proposed.Firstly,the convolution neural network is used to extract the image information supporting the dataset,and the image representation is obtained.Then the manifold structure of data features and the fiber bundle of data features are constructed.The input query set selects the manifold section of the current new task to obtain the fiber suitable for the new task,so as to get the correct label of the image.Experimental results show that the model based on the proposed algorithm(FBBML)achieves the best accuracy performance compared with the standard four-layer convolutional neural network model on the common data set(mini-ImageNet).At the same time,the fiber bundle theory is introduced into meta learning,which makes the algorithm more interpretable.

关 键 词：贝叶斯 流形 纤维丛 分类 卷积神经网络 元学习 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]