LGFeb 22, 2017
Training a Subsampling Mechanism in Expectation
arXiv:1702.06914v32 citations
Originality Synthesis-oriented
AI Analysis
This work addresses a methodological challenge in machine learning for researchers, but it is incremental as it focuses on a specific mechanism with limited testing.
The authors tackled the problem of training a subsampling mechanism for sequences by computing its expected output to enable standard backpropagation, and they tested it on a toy problem while noting its shortcomings.
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.