where y(Cn Pi) is the output of the network for class Cn on image patch Pi, M is the number of patches in the im- age, and k 1 is a parameter. Higher values of k focus on the highest scoring patches and attenuate the contributions of low- and mid-scoring patches. The value of k = 5 was optimized on the validation set and is fixed in our experi- ments.Note that patch scores could be computed much more efficiently by performing large convolutions on adequately subsampled versions of the full image, as described for in- stance in [12]. This would permit a denser patch coverage at a lower computation cost.