摘要
卷积神经网络是对于人脑的高度抽象,它是深度学习的重要组成部分。对于卷积神经网络的研究,一方面有助于更准确地进行图像的分类与识别,另一方面,有助于人类更真实地模拟人脑,为人工智能的发展指明了方向。分析比较了Sigmoid、Tanh、Re Lu、Softplus4种激活函数的优缺点。结合Re Lu和Softplus两种激活函数的优点,设计并构造了一种分段激活函数。最后,基于Theano框架和这5种激活函数,分别构建了5种卷积神经网络,并对Cifar-10数据集进行了分类识别。实验结果表明,基于改进后的激活函数所构造的卷积神经网络,不仅收敛速度更快,而且可以更加有效地提高分类的准确率。
Convolutional neural network is a high degree of abstraction to the human brain and an important part of deep learning. For re- search on it, on the one hand, it is helpful for a more accurate image classification and recognition. On the other hand,the human brain can be more truly simulated, which points out the direction for the development of artificial intelligence. First the advantages and disadvanta- ges of four kinds of activation functions such as Sigmoid ,Tanh ,ReLu and Softplus are analyzed and compared. Then ,combined with the advantages of ReLu and Softplus, a piecewise activation function is designed and constructed. Finally, based on Theano framework and these activation functions, five convolutional neural networks are established respectively for classification recognition on the Cifar-10 da- m sets. The experimental results show that the convolution neural network based on the improved activation function not only converges faster, but also improves the classification accuracy more effectively.
出处
《计算机技术与发展》
2017年第12期77-80,共4页
Computer Technology and Development
基金
国家自然科学基金资助项目(61271082)
江苏省重点研发计划(BE2015700)
江苏省自然科学基金(BK20141432)
关键词
卷积神经网络
深度学习
人工智能
激活函数
convolutional neural network
deep learning
artificial intelligence
activation function
