[dontreact]

AllConv networks perform better and are an even simpler modification of the typical CNN architecture: just replace max pool layers with strides convolution. Why did the authors not benchmark against state of the art (also see fractional max pooling)

https://arxiv.org/abs/1412.6806

[ironislands]

Thanks for pointing me to this reference. I have implemented the baseline model mentioned in this paper. A sharp rise of 4% in classification accuracy is clearly observable. While there would be an improvement due to theoretical reasons, the difference between improved accuracy and original accuracy might get less on introducing other factors. It is on my TODO to add new results on different models and by introducing other things in a simple CNN as well.