Backpropagation through Combinatorial Algorithms: Identity with Projection Works

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Sahoo, Subham Sekhar, Paulus, Anselm, Vlastelica, Marin, Musil, Vít, Kuleshov, Volodymyr, Martius, Georg
Format: Preprint
Published: 2022
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915061982822400
author Sahoo, Subham Sekhar
Paulus, Anselm
Vlastelica, Marin
Musil, Vít
Kuleshov, Volodymyr
Martius, Georg
author_facet Sahoo, Subham Sekhar
Paulus, Anselm
Vlastelica, Marin
Musil, Vít
Kuleshov, Volodymyr
Martius, Georg
contents Embedding discrete solvers as differentiable layers has given modern deep learning architectures combinatorial expressivity and discrete reasoning capabilities. The derivative of these solvers is zero or undefined, therefore a meaningful replacement is crucial for effective gradient-based learning. Prior works rely on smoothing the solver with input perturbations, relaxing the solver to continuous problems, or interpolating the loss landscape with techniques that typically require additional solver calls, introduce extra hyper-parameters, or compromise performance. We propose a principled approach to exploit the geometry of the discrete solution space to treat the solver as a negative identity on the backward pass and further provide a theoretical justification. Our experiments demonstrate that such a straightforward hyper-parameter-free approach is able to compete with previous more complex methods on numerous experiments such as backpropagation through discrete samplers, deep graph matching, and image retrieval. Furthermore, we substitute the previously proposed problem-specific and label-dependent margin with a generic regularization procedure that prevents cost collapse and increases robustness.
format Preprint
id arxiv_https___arxiv_org_abs_2205_15213
institution arXiv
publishDate 2022
record_format arxiv
spellingShingle Backpropagation through Combinatorial Algorithms: Identity with Projection Works
Sahoo, Subham Sekhar
Paulus, Anselm
Vlastelica, Marin
Musil, Vít
Kuleshov, Volodymyr
Martius, Georg
Machine Learning
Embedding discrete solvers as differentiable layers has given modern deep learning architectures combinatorial expressivity and discrete reasoning capabilities. The derivative of these solvers is zero or undefined, therefore a meaningful replacement is crucial for effective gradient-based learning. Prior works rely on smoothing the solver with input perturbations, relaxing the solver to continuous problems, or interpolating the loss landscape with techniques that typically require additional solver calls, introduce extra hyper-parameters, or compromise performance. We propose a principled approach to exploit the geometry of the discrete solution space to treat the solver as a negative identity on the backward pass and further provide a theoretical justification. Our experiments demonstrate that such a straightforward hyper-parameter-free approach is able to compete with previous more complex methods on numerous experiments such as backpropagation through discrete samplers, deep graph matching, and image retrieval. Furthermore, we substitute the previously proposed problem-specific and label-dependent margin with a generic regularization procedure that prevents cost collapse and increases robustness.
title Backpropagation through Combinatorial Algorithms: Identity with Projection Works
topic Machine Learning
url https://arxiv.org/abs/2205.15213