Back to search

Georgia Institute of Technology

Overcoming Noise and Variations In Low-Precision Neural Networks

Abstract

dc:description.abstract

This work explores the impact of various design and training choices on the resilience of a neural network when subjected to noise and/or device variations. Simulations were performed under the expectation that the neural network would be implemented on analog hardware; this context asserts that there will be random noise within the circuit as well as variations in device characteristics between each fabricated device. The results show how noise can be added during the training process to reduce the impact of post-training noise. Architectural choices for the neural network also directly impact the performance variation between devices. The simulated neural networks were more robust to noise with a minimal architecture with fewer layers; if more neurons are needed for better fitting, networks with more neurons in shallow layers and fewer in deeper layers closer to the output tend to perform better. The paper also demonstrates that activation functions with lower slopes do a better job of suppressing noise in the neural network. It also shown that the accuracy can be made more consistent by introducing sparsity into the neural network. To that end, an evaluation is included of different methods for generating sparse architectures for smaller neural networks. A new method is proposed that consistently outperforms the most common methods used in larger, deeper networks.

Degree

thesis:*
Level thesis:degree_level
Doctoral
Department dc:contributor.department
Electrical and Computer Engineering
Grantor dc:publisher
Georgia Institute of Technology
Year dc:date.issued
2021

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Janke, Devon D.
Advisor dc:contributor.advisor
  • Anderson, David V.
Committee members dc:contributor.committeemember
  • Raychowdhury, Arijit
  • Lanterman, Aaron
  • Hyesoon, Kim
  • Li, Shaolan

Subjects

dc:subject × 7

Rights

Language dc:language.iso
en_US

Identifiers

dc:identifier.*
Handle dc:identifier.uri
http://hdl.handle.net/1853/65025

Chain of custody

source
Harvested from
Georgia Tech
Base URL
repository.gatech.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Janke, Devon D.. Overcoming Noise and Variations In Low-Precision Neural Networks. Doctoral thesis, Georgia Institute of Technology, 2021. http://hdl.handle.net/1853/65025