0

Example taken from VGG16 architecture

enter image description here

kwquan
  • 16
  • 2

1 Answers1

1

VGG was among the first networks that were developed in the context of deep CNNs. The number you see, which stands for the number of filters, which you already know what they mean, were chosen to increase in number as the network increased in depth.

One of the intuitions for this was that as the network increased in depth, more complex features could be learned, and in turn the complexity of the features could have been only captured by more filters. This architecture type is not present only in VGG, but also in later architectures, which as they gradually increase the depth, the number of filters many times increase by a factor of *2.

Timbus Calin
  • 13,809
  • 5
  • 41
  • 59
  • The question is: why 8,16,32,64 ... Why not 4,8,12,24 ...? – kwquan Oct 18 '21 at 09:52
  • Perhaps 4 filters is a too little number to start. Such details can only be answered by the authors of the paper. I explained to you the logic of the architectures, I don't think the discussion should be focused on why 8 and not 4, but the overall tendency&progress. – Timbus Calin Oct 18 '21 at 11:00
  • Alright. Thanks a lot – kwquan Oct 18 '21 at 11:29
  • If it helped you solve your doubts, it is customary to accept the answer/upvote it. – Timbus Calin Oct 18 '21 at 11:33