1
$\begingroup$

When I extract my features from my CNN, it doesn't look like this:

This

And those pictures are not just representation. From this article it can be seen that these features are actual extracted features from real CNN.

However, the features that I extracted look exactly like this:

This

Why is this representation not like the first picture? Is it the correct one?

$\endgroup$

1 Answer 1

2
$\begingroup$

I don't think you are comparing like with like.

In the left-most panel of the first image, you are seeing the weights in each kernel (one channel from one convolutional layer). These are yellow in the figure below. The size of the kernels is determined by the hyperparameters of the network; they might have a size like 3 × 3 or 31 × 31. I'm not 100% certain about the other two panels; the right-most looks more like convolutional products than filters.

In the second, you are certainly looking at the activations when given a particular input example. These images are the result of convolving the input with the kernels; this part is pink in the figure below. Their size depends on the input image size, the kernel size, and the convolution parameters.

From the article you linked to:

Cartoon of image convolution

$\endgroup$
2
  • $\begingroup$ But people are not saying those are kernels. Those are feature maps. google.com/… $\endgroup$
    – canP
    Commented Jul 8, 2022 at 11:28
  • 2
    $\begingroup$ I agree it's likely the right-most panel in the first figure you posted shows feature maps, but I'm pretty sure the elements of the first panel, and I believe the second as well, are the learned filters. $\endgroup$
    – kwinkunks
    Commented Jul 8, 2022 at 11:52

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.