摘要
在当前语音识别和图像识别领域,卷积神经网络已经取得了很大的成功。现有的Lenet-5卷积神经网络是多层网络结构,但是大量实验表明,从全链接层到输入层的回调影响了最终的精度,特别是在有限数据量的情况下。因此提出了单层回调的Lenet-5算法,即在Lenet-5卷积神经网络的卷积层后添加一个临时输出层,与真实标签进行比较,根据误差函数对层间参数进行回调,并用全球手写数字MNIST数据集进行训练和测试。实验表明,即使在有限数据量的情况下算法的精度仍能得到提高。
In the area of speech recognition and image recognition, convolutional neural network has achieved a great success, the structure of Lenet-5 convolutional neural network is multilayer, but a lot of experiments show that the callback from full connection layer to the input layer influences the accuracy, especially in the case of limited data sets. So, the single layer callback Lenet-5 algorithm is proposed, which adds a temporary output layer in the end of each convolutional layer, compares the output with the real labels, adjusts the parameter of each layer according to the error function, and uses the global MNIST data base to train and test. The experiment proves that the accuracy of the algorithm can be improved even in case of limited data sets.
出处
《计算机时代》
2016年第8期4-6,12,共4页
Computer Era
关键词
卷积神经网络
Lenet-5
全链接
卷积
池化
回调
精度
convolutional neural network
Lenet-5
full-connection
convolution
pooling
callback
accuracy
