Training a Subsampling Mechanism in Expectation
Abstract
We describe a mechanism for subsampling sequences and show how to compute its expected output so that it can be trained with standard backpropagation. We test this approach on a simple toy problem and discuss its shortcomings.
Research Areas
