Skip to main content

The Seeing Stack

This video presents the same text shown beside it, spoken and on screen. It adds nothing the text does not say.

State

A convolutional network stacks template maps feeding template maps, with bends and coarsening between — each small neighborhood's strongest response kept, the rest shed.

Show

Stacked layers grow each output's field of influence until a final response can depend on the whole image; competition-era stacks ran past 150 layers.

Watch for

The coarsening buys tolerance to small shifts — and spends fine position information to get it.