Click on the top " AI proper way ", select the "star" Public No.
Heavy dry goods, the first time served
Convolution neural network as early as in the 1980s, he was dean of the neural network Lecun proposed [LeNet-5, LeCun 1980], but because at that time the amount of data, computing power and other issues, have not been widely used.
Inspired convolution neural network from the 1950s Nobel Prize for biology. Hubel & Wiesel, 1959, cat brain regions responsible for vision, for a small segment of interest (rather than fish), this area is far away from the eyes (and not near).
Convolutional Layer 1. Layer convolution
Layer has a convolution filter, continue to slide in image space, dot for dot
The results of dot filter is a numerical value, after sliding dot, it will form a smaller picture than the original layer thickness of 1
filter can be many, we are an example to six filter
2. convolution process Convolutional Process
The process of convolution, convolution is a series of layers in combination:
Each layer extracted feature from the low lever, mid level, high level can be classified into levels
Here is a picture of the classification process:
3. convolution calculation Convolutional Compute
(1) to calculate the activation map size 7 * 7 size image, for example
(2)zero padding
zero padding can keep the size of the picture convolution
(3) Calculation
Example: input picture is 32x32x3 th filter 10 has a 5x5, stride is 1, pad 2
Size of the output? How many parameters?
Output volume size: (32+2*2-5)/1+1 = 32 spatially, so 32x32x10
Number of parameters in this layer? each filter has 553 + 1 = 76 params => 76*10 = 760
(4) summary
本专栏图片、公式很多来自台湾大学李弘毅老师、斯坦福大学cs229、cs231n 、斯坦福大学cs224n课程。在这里,感谢这些经典课程,向他们致敬!
作者简介:武强 兰州大学博士,谷歌全球开发专家Google Develop Expert(GDE Machine Learing 方向)
CSDN:https://me.csdn.net/dukuku5038
知乎:https://www.zhihu.com/people/Dr.Wu/activities
漫画人工智能公众号:DayuAI-Founder
系列笔记: