In ALEXNET, I dont understand why last convolution layer has 128 * 2 filters

#1

The understanding that I got from the webinar on CNN was that the later part of the network learns complex features and thus require more number of filters. But in ALEXNET the layers have
48 * 2
128 * 2
192 * 2
192 * 2
And then finally : " 128 * 2 "
Is there any particular explanation for this?

#2

Hey @Lakshita, yes your understanding is truly apt, but you may infer that, sometimes in models, we first increases the filters count, and towards the last layer we ten to decrease filter size. This is because if we directly flatten() than the number of parameters may be much more. So we decrease the filter size so that after adding flatten() layere we introduce dense layers, so number of parameters may not be enormously high.

Hope this resolved your doubt.
Plz mark the doubt as resolved in my doubts section. :blush: