Changing the resolution of the CNN to do RAM

[emoen]

In the paper chapter 4.5.1 Network settings for RAM: it is mentioned that the network is changed to produce larger RAM images - 54x54. How is then the RAM (of the changed network) comparable to the results of the original network (e.g net-5)?

[emoen]

If you have a network you dont want to modify - like say - InceptionV4 where the input is (299,299) and the output width is (8,8) how can you increase the resolution without removing alot of layers, and therefore change the architecture significantly? Can you insert GAP layers in the middle of the network to produce RAMs? What should the weights, w, on the conv layer to GAP be?