Adaptive Neural Network Architectures for Power Aware Inference

Skyler Anderson, Nagadastagiri Challapalle, John Sampson, Vijaykrishnan Narayanan

Research output: Contribution to journalArticle

Abstract

As an increasingly diverse array of edge devices become platforms for neural networks, the power and compute limitations of these platforms become key design constraints, often imposing trade-offs among performance, accuracy, and power/energy requirements for inference tasks. Moreover, in edge deployment scenarios, dynamic variability in both the power and computation capabilities of these platforms is an equally important driver of design decisions in matching a neural network model with a given end device and deployment scenario. In this work we propose a novel method for adaptive neural network architectures to boost performance as power becomes available without necessitating a complete reconfiguration of model parameters. We show that, due to the stochastic nature of neural networks, models can be constructed with enough independence to be effectively ensembled while sharing a common convolutional base. This allows for a trade-off between power consumption and model accuracy without redeploying an entirely new model. Our method is agnostic to the base network and can be added to existing networks with minimal retraining. We find this approach particularly well suited for IoT devices and distributed autonomous applications where available power and compute resources are time varying.

Original languageEnglish (US)
JournalIEEE Design and Test
DOIs
StateAccepted/In press - Jan 1 2019

Fingerprint

Network architecture
Neural networks
Electric power utilization

All Science Journal Classification (ASJC) codes

  • Software
  • Hardware and Architecture
  • Electrical and Electronic Engineering

Cite this

Anderson, Skyler ; Challapalle, Nagadastagiri ; Sampson, John ; Narayanan, Vijaykrishnan. / Adaptive Neural Network Architectures for Power Aware Inference. In: IEEE Design and Test. 2019.
@article{128a439436794d8682cad0f2b60b83a6,
title = "Adaptive Neural Network Architectures for Power Aware Inference",
abstract = "As an increasingly diverse array of edge devices become platforms for neural networks, the power and compute limitations of these platforms become key design constraints, often imposing trade-offs among performance, accuracy, and power/energy requirements for inference tasks. Moreover, in edge deployment scenarios, dynamic variability in both the power and computation capabilities of these platforms is an equally important driver of design decisions in matching a neural network model with a given end device and deployment scenario. In this work we propose a novel method for adaptive neural network architectures to boost performance as power becomes available without necessitating a complete reconfiguration of model parameters. We show that, due to the stochastic nature of neural networks, models can be constructed with enough independence to be effectively ensembled while sharing a common convolutional base. This allows for a trade-off between power consumption and model accuracy without redeploying an entirely new model. Our method is agnostic to the base network and can be added to existing networks with minimal retraining. We find this approach particularly well suited for IoT devices and distributed autonomous applications where available power and compute resources are time varying.",
author = "Skyler Anderson and Nagadastagiri Challapalle and John Sampson and Vijaykrishnan Narayanan",
year = "2019",
month = "1",
day = "1",
doi = "10.1109/MDAT.2019.2947258",
language = "English (US)",
journal = "IEEE Design and Test",
issn = "2168-2356",
publisher = "IEEE Computer Society",

}

Adaptive Neural Network Architectures for Power Aware Inference. / Anderson, Skyler; Challapalle, Nagadastagiri; Sampson, John; Narayanan, Vijaykrishnan.

In: IEEE Design and Test, 01.01.2019.

Research output: Contribution to journalArticle

TY - JOUR

T1 - Adaptive Neural Network Architectures for Power Aware Inference

AU - Anderson, Skyler

AU - Challapalle, Nagadastagiri

AU - Sampson, John

AU - Narayanan, Vijaykrishnan

PY - 2019/1/1

Y1 - 2019/1/1

N2 - As an increasingly diverse array of edge devices become platforms for neural networks, the power and compute limitations of these platforms become key design constraints, often imposing trade-offs among performance, accuracy, and power/energy requirements for inference tasks. Moreover, in edge deployment scenarios, dynamic variability in both the power and computation capabilities of these platforms is an equally important driver of design decisions in matching a neural network model with a given end device and deployment scenario. In this work we propose a novel method for adaptive neural network architectures to boost performance as power becomes available without necessitating a complete reconfiguration of model parameters. We show that, due to the stochastic nature of neural networks, models can be constructed with enough independence to be effectively ensembled while sharing a common convolutional base. This allows for a trade-off between power consumption and model accuracy without redeploying an entirely new model. Our method is agnostic to the base network and can be added to existing networks with minimal retraining. We find this approach particularly well suited for IoT devices and distributed autonomous applications where available power and compute resources are time varying.

AB - As an increasingly diverse array of edge devices become platforms for neural networks, the power and compute limitations of these platforms become key design constraints, often imposing trade-offs among performance, accuracy, and power/energy requirements for inference tasks. Moreover, in edge deployment scenarios, dynamic variability in both the power and computation capabilities of these platforms is an equally important driver of design decisions in matching a neural network model with a given end device and deployment scenario. In this work we propose a novel method for adaptive neural network architectures to boost performance as power becomes available without necessitating a complete reconfiguration of model parameters. We show that, due to the stochastic nature of neural networks, models can be constructed with enough independence to be effectively ensembled while sharing a common convolutional base. This allows for a trade-off between power consumption and model accuracy without redeploying an entirely new model. Our method is agnostic to the base network and can be added to existing networks with minimal retraining. We find this approach particularly well suited for IoT devices and distributed autonomous applications where available power and compute resources are time varying.

UR - http://www.scopus.com/inward/record.url?scp=85073511749&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85073511749&partnerID=8YFLogxK

U2 - 10.1109/MDAT.2019.2947258

DO - 10.1109/MDAT.2019.2947258

M3 - Article

AN - SCOPUS:85073511749

JO - IEEE Design and Test

JF - IEEE Design and Test

SN - 2168-2356

ER -