Saved in:
Bibliographic Details
Main Authors: Zhang, Yibo Jacky, Tang, Zeyu, Koyejo, Sanmi
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.27946
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910264594530304
author Zhang, Yibo Jacky
Tang, Zeyu
Koyejo, Sanmi
author_facet Zhang, Yibo Jacky
Tang, Zeyu
Koyejo, Sanmi
contents Backpropagation is the default learning rule for artificial neural networks and is often treated as the settled approach whenever differentiability is available. In this work, we revisit this convention through a theoretical lens of sample efficiency. We introduce a unified vectorized feedback framework for loss-based and reward-based learning on computational graphs, in which synthetic gradients emerge as a natural alternative to backpropagation. We characterize the conditions under which synthetic gradients can achieve a lower gradient-estimation mean squared error than backpropagation. We construct examples illustrating that this sample efficiency advantage can be arbitrarily large. Experiments on contextual bandits and reinforcement learning tasks demonstrate the potential of our theoretical findings.
format Preprint
id arxiv_https___arxiv_org_abs_2605_27946
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Is Backpropagation Optimal? When Synthetic Gradients Improve Sample Efficiency
Zhang, Yibo Jacky
Tang, Zeyu
Koyejo, Sanmi
Machine Learning
Backpropagation is the default learning rule for artificial neural networks and is often treated as the settled approach whenever differentiability is available. In this work, we revisit this convention through a theoretical lens of sample efficiency. We introduce a unified vectorized feedback framework for loss-based and reward-based learning on computational graphs, in which synthetic gradients emerge as a natural alternative to backpropagation. We characterize the conditions under which synthetic gradients can achieve a lower gradient-estimation mean squared error than backpropagation. We construct examples illustrating that this sample efficiency advantage can be arbitrarily large. Experiments on contextual bandits and reinforcement learning tasks demonstrate the potential of our theoretical findings.
title Is Backpropagation Optimal? When Synthetic Gradients Improve Sample Efficiency
topic Machine Learning
url https://arxiv.org/abs/2605.27946