What's the guideline to design a CNN

I am on the last week of C4. I keep wondering how to design or understand a CNN? Like how deep the network should be for a specific task, what’s the meaning of each hidden layers? Why we make dimension smaller but channel bigger?

When the course says the first layer maybe to detect the "edge"s of the image Does it really design in this way or just guess? Or the way to find a good arch for a CNN is just to keep trying different combination of parameters? That seems too random (I know we can start from an existing work and do tunning based on that). Not sure if I make sense here.

Hello @Jack_Tao,

On first order, we balance between over-fitting and under-fitting.

We don’t keep spatial information at the end, and only want features (channels) to do classification.

No, it wasn’t by manual design. It was trained to be. Early layer is only able to detect simple features because its filters only capture a smaller area of the image.

Cheers,
Raymond

1 Like