AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】

论文题目和原文地址:ImageNet Classification with Deep Convolutional Neural Networksvolu

卷积在三通道上的操作

AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
一般来说有几个卷积核,最终得到的featuremap就有几个。三通道的图片对应的卷积核也是三通道的,三个通道分别卷积之后对应位置相加填在featuremap中。比如上图中最右边的3就是通过三个通道卷积之后得到的。
池化
也叫做下采样,pooling。有平均池化和最大值池化,可以减少featuremap的尺寸,减少计算量,避免过拟合,引入平移不变性。
Relu不饱和的**函数可以解决梯度消失问题。
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
在下面的网络中,有5个卷积层和3个全连接层,所以有8个能够自己学习权重的层。
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
提出了局部响应归一化,就使得不同通道上的同一像素位置不会出现太多的高**,而且会抑制附近的**
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
重叠池化
在这里插入图片描述
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
水平翻转,随机裁剪,平移变换,颜色、光照变换增大数据集,减少过拟合
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
减少过拟合之dropout随机掐死一些神经元。减少了神经元的存在依赖适应性,逼着神经元和不同的神经元进行合作。
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】

论文阅读

AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
卷积网络中神经元的个数=featuremap的元素个数=卷积的次数

AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】
AlexNet论文阅读【ImageNet Classification with Deep Conventional Neural Networks】