深度学习中先验知识的应用  

Application of prior knowledge in deep learning

作　　者：宣冬梅[1] 王菊韵[2] 于华[1] 赵佳[1] 

机构地区：[1]中国科学院大学工程管理与信息技术学院,北京100049 [2]中国传媒大学理学院,北京100024

出　　处：《计算机工程与设计》2015年第11期3087-3091,共5页Computer Engineering and Design

基　　金：国家重点基础研究发展计划基金项目(2011CB706900);国家自然科学基金项目(70971128);北京市自然科学基金项目(9102022)

摘　　要：在多分类问题中使用深度学习方法提高分类效果的模型通常很复杂,且运行时间较长。为解决该问题,提出先验知识与深度学习结合的方法。对深度学习中先验知识在多分类问题中的应用进行研究,得到3个结果:当先验知识较准确时,带有先验的分类方法能大大提高分类效果;神经元之间的连接权重绝对值将减小,模型能够求解;深度神经网络中使用的网络结构会更简单。在MNIST手写体数据集上进行的实验验证了该方法的有效性和其结果的准确性。The model is usually very complex and time-consuming when deep learning method is used to improve the result of classification in the multi-classification.To solve the problem,the method that combined prior knowledge and deep learning method was put forward.The application of prior knowledge in multi-classification problems was studied in the context of deep learning and three results were obtained.The identification rate was improved when the prior knowledge was accurate;the absolute value of link weights between neurons was smaller and the model was solved;the simpler neural network architecture was constructed with prior knowledge.Results of experiment on MNIST verify the validity and efficiency of the presented method.

关 键 词：分类问题 深度学习 先验知识 权重 手写体数据集 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]