Given that we have a fixed piece of hardware to use on networks of ever-increasing size, it takes more power to run the network, because we have to reuse the same ALUs multiple times: https://twitter.com/ylecun/status/1469690849357410306.
There are lots of challenges involved in efficiency because of the number of filters, trying to maintain numerical stability while using half/single-precision, etc., but using repetitive filters can allow for optimizations that can be seen in VGGNet.
Given that we have a fixed piece of hardware to use on networks of ever-increasing size, it takes more power to run the network, because we have to reuse the same ALUs multiple times: https://twitter.com/ylecun/status/1469690849357410306.
There are lots of challenges involved in efficiency because of the number of filters, trying to maintain numerical stability while using half/single-precision, etc., but using repetitive filters can allow for optimizations that can be seen in VGGNet.