Problem regarding the weights and biases in CNN

Question

I am currently learning about CNN.

Till now I have understood the following:

1. We pass the image as an input.
2. In the 1st Convolution layer, we apply all the filters to the 2d image and apply an activation function, max pool the image to reduce the dimensionality and pass the output to the next layer.
3. In the next layer, we repeat the process and pass the output to a Fully connected layer.
4. This layer then acts as a basic neural network and predicts the outcome.

I understand about the weights and biases in the Fully Connected layers in the later stages but I am not able to understand how do we assign weights to the initial Convolution layers in the CNN?

Do we assign weights only to the Fully Connected Layers? Or do we have to assign weights to Convolution Layer in a particular fashion?

Answer 1

Alright, first, I encourage you to keep reading about CNNs!

I will try to explain what it is actually doing. CNNs are usually divided into two parts. The first part is made up of convolutional layers and poolings. These act as the eyes and learn to see different things. The weights of the conv layers are the window's values that are slided through the inputs, they have to be initialized just as the weights of a fully connected layer, same with the bias. The second part of the CNNs are the fully connected layers added in the end. These act as the brain, and use what the 'eyes' are seeing for classification.

The new weights and bias of the filters are calculated with backpropagation the same as if it were a fully connected network.

I hope I made it a little more clear. If you have more questions feel free to ask.

Cheers.