This is a bit of a riff on hypernetworks and Multiple Object Recognition with Visual Attention. Just testing the information efficiency of a model which queries for sparse samples of processed information rather than taking everything in a large processing stack.