2. Fundamentals of CNNs and RNNs

This is a note from coursera lecture : https://www.coursera.org/learn/cnns-and-rnns

 

• The final result after convolution and pooling is feature map.

 

• How many numbers in ther filter needed?

Number of input channels x filtersize x Number of output channels.

Assuming that filter size is 3 x 3.

The number of numbers in the filter that I need to fill is n x 3 x 3 x m.

 

For example,

if the size of input image is 32 x 32 x 3(height byte, width byte, num of channels),

the size of filter is 3 x 3,

the number of filter's channel is 3, same as the number of input's channel,

the number of filter is 64(same as number of output channel).

Therefore, filter parameter in this case is 3 x 3 x 3 x 64.

 

• Pooling does not change the number of channels.

The number of channels remains same. Only the size of the image reduces.

 

• Who determines filters?

CNN can be converted into neural networks, in other words, convolution filters are converted into connection weights. During training FCNN, filters are found by the gradient descent methods.

Kernel can be converted to connection weight of neural network.

 

• Convolution and pooling can be converted to Neural networks layers.

For example, in convolution operation, input and output image can be converted to neural network layer.

Each pixel changes to node in neural network layer.

 

And in (max) pooling operation, the result of pooling can be converted to neural network layer.

 

In other words,

Convolution + Threshold + Pooling can be converted to Neural Network.

↓

 

Let's look at the overview.

More channels, more nodes.

↓

Convolution layers and pooling layers are repeated. 

Convolution layers are mandatory but pooling layers are optional.