2
$\begingroup$

Context: I'm building a CNN on MATLAB to classify wallpaper groups.

I'm using the following network type.

CONV -> ReLU -> POOL -> CONV -> ReLU -> POOL -> FC -> DROPOUT -> FC -> SOFTMAX

Should the parameters for the first CONV or POOL layers be the same as the second (or later) CONV or POOL layers succeeding it? If not, is there a meaningful way to choose parameters for the repeated succeeding layers?

For example, for the first CONV layer you may pick a 5x5 filter. Perhaps for the second CONV layer it's more meaningful to pick a certain different size filter?

$\endgroup$

1 Answer 1

1
$\begingroup$

Finding an appropriate architecture is somehow practical. Those hyper-parameters you are talking may be different. Try to use a base architecture and then train your model. If it does not learn your model try to change the hyper parameters. It is an iterative operation to find a good model. There are a lot of debates, but there is not exactly consensus for making good models in CNNs. You can think of if your model does not learn your data, it needs more features to be learned by the CNN. The behavior of the layers are somehow as follows:

Max Pooling

It is used for adding spacial invariance to the inputs of its layer. It is also used for decreasing the input size. By increasing its size, the two mentioned behavior would increase.

Convoluional Layers

These layers are used for extracting features to reduce the cost function in order to learn data. If you increase the size of these layers they behavior would be more vast, they would find features of a larger region.

Take a look at here and here.

$\endgroup$
0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.