to sum up:
全连接侧重 特征的精确位置
卷积层侧重 特征的相对位置
Thoughts: Each train can learn the same local characteristics do, and why
Convolution kernel
- A convolution kernel is a local feature
- Therefore, output of the convolution layer may depend less precise location features (compared to the fully connected layers)
Why do you say:
Because the right to fully connected layers of heavy and location are closely related
but closely associated with the right weight and features a convolution layer itself
- Advantages: small amount of calculation, high generalization
Understand the internal logic
Ikekakaku
- A filter
- Role: to make features more focused
One more convolution kernel used what's the use:
1.提取出更多的局部特征
2.减轻池化层的信息损失
data
On a sample image (upsampling) and downsampling (subsampled)
typically convolutional neural network architecture