It seems like you use the same skeleton with the NIN model bu the supervised layers. And any number is same like channel numbers and filter sizes. Even you introduces more params with supervised layers. however your paper states that number of parameters are arranged to keep the values comparable with the original NIN model.
It seems like you use the same skeleton with the NIN model bu the supervised layers. And any number is same like channel numbers and filter sizes. Even you introduces more params with supervised layers. however your paper states that number of parameters are arranged to keep the values comparable with the original NIN model.