I have noticed, as a trend, people seem to "taper" the size of their filters as a convolutional network progresses. By this I mean they begin convolving the image/patch with a larger filter, and slowly decrease the size each layer until the output.

I have also employed this approach and had very good results, however I am not sure why.

Is there a reason this should be done, or is it just "black magic"?



The receptive fields of filters projected back on the input is exponentially increasing in each layer. This method, I think, is one way to keep RFs in under control and make sure that hierarchy is not broken, i.e., RFs are slowly increasing.

This is not a must, though, I recommend to read the following where authors propose to use only 3x3 filters.

Simonyan, Karen, and Andrew Zisserman. "Very deep convolutional networks for large-scale image recognition." arXiv preprint arXiv:1409.1556 (2014).

  • $\begingroup$ thanks! I am also wondering about the number of feature maps attributed to each layer. something that seems to commonly increase over the number of layers, however there is little discussion around this (unless i am not looking hard enough :-D ) $\endgroup$
    – JB1
    Jul 14 '16 at 10:57

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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