Convolutional Layers

Example of a typical convolutional network.

Let us take an example of a network with input image having dimensions 39 x 39 x 3. So as per the notation we get $n_H^{[0]}=n_W^{[0]}=39$ and $n_C^{[0]}=3$.

For layer 1 we will take $f^{[1]}=3$, $s^{[1]}=1$ and $p^{[1]}=0$. Suppose we use 10 filters in this layer. Each filter has dimension $[3,3,3]$.

So after the convolution we get the resultant image having dimensions 37 x 37 x 10. So we have $n_H^{[1]}=n_W^{[1]}=37$ and $n_C^{[1]}=10$.

For layer 2 we will take $f^{[2]}=5$, $s^{[2]}=2$, $p^{[2]}=0$ and $n_C^{[2]}=20$. So after layer 2 we will get an resultant image having dimension 17 x 17 x 20. So we have $n_H^{[2]}=n_W^{[2]}=17$.

For layer 3 let us use $f^{[3]}=5$, $s^{[3]}=2$ $p^{[3]}=0$ and $n_C^{[3]}=40$. So after layer 3 we get the image having dimensions 7 x 7 x 40.

After this layer we take all the $7\times 7\times 40=1960$ values and flatten them into a vector and then feed this to a logistic/softmax unit depending upon the type of classification we are performing.

A typical convolutional network has 3 types of layers

• Convolutional layer (CONV)
• Pooling layer (POOL)
• Fully connected layer (FC)

See now all this connect to Convolutional Neural Networks