Reading through this paper, this is a logical extension of the recent work published by FAIR (Facebook AI Research) that proved that with careful implementation

Reading through this paper, this is a logical extension of the recent work published by FAIR (Facebook AI Research) that proved that with careful implementation of FFT convolution that the speed up is very significant. The Facebook work though still had all the learning happening in the spatial not frequency domain. They are doing all the updating in the frequency domain, and have introduced a new type of pooling that uses stochastic resolution reduction in the frequency domain that seems very useful. Very interesting paper, I'm keen to try out the techniques myself.