Please enable JavaScript.
Coggle requires JavaScript to display documents.
convolutional networks - Coggle Diagram
convolutional networks
傳統NN的限制，CNN可以做到
not translation-invariant
convolution layer
max pooling
not rotation-invariant
not scale-invariant
用卷積代替矩陣乘法
maximum likelihood
back-propagation
trainable parameter
channels 3
kernels 16
feature maps 16
=(3
*3)
3*16+16
Densely connected vs convolution layer
Densely connected
global patterns
MNIST digit
convolution layer
local patterns
image can be edges texture...etc
translation invariant.
spatial hierarchies of patterns
spatial hierarchy of visual
modules
貓
Deconvolution
類似於矩陣轉至的乘法
從輸出到輸入反向傳播
特性
稀疏連接
receptive field
跟下一層有關係的節點，layer大field大
參數共享
常用於圖像影片
max pooling
cross channel pooling
pooling function
the average of a rectangular neighborhood
the L2 norm of a rectangular neighborhood
a weighted average based on the distance from the central pixel
with downsampling
size 縮小
另一個方法是convolution with stride
zero padding