What are the trainable parameters of convolutional layers?

there is this thig that puzzles me.
Do convolutional layers have trainable parameters (weights)? If so, what are they?
From the lectures I understand that convolutional layers “just” perform some filtering on their input, but I also understand that the values on the filter matrix are predetermined. Then I can’t understand what it means to train the convolutional part on a network.

Thank you!


This is a nice video from Prof Ng from Deep Leanring Specialization explaining your question.

Convolutional Layer. The values of the conv filter are not determined but to be learned as weights.

1 Like