A new kind of pooling layer for faster and sharper convergence
Sahil Singla

I would like to see a comparison between this and vanilla softmax pooling, suppose a_1, a_2, a_3, a_4 are your pooled activations. Apply a linear function x → wx, where w is a parameter to be learned. Then take the softmax of the result. This specialises to max pooling for large w and average pooling for small w, but doesn’t require any reordering of elements.

One clap, two clap, three clap, forty?

By clapping more or less, you can signal to us which stories really stand out.