Link to the University of Pittsburgh Homepage
Link to the University Library System Homepage Link to the Contact Us Form

A Case Study in Practical Neuromorphic Computing: Heartbeat Classification on the Loihi Neuromorphic Processor

Buettner, Kyle (2021) A Case Study in Practical Neuromorphic Computing: Heartbeat Classification on the Loihi Neuromorphic Processor. Master's Thesis, University of Pittsburgh. (Unpublished)

[img] PDF
Restricted to University of Pittsburgh users only until 13 June 2022.

Download (913kB) | Request a Copy

Abstract

One potential method of efficiently deploying deep neural networks is with neuromorphic computing, a paradigm of processing that emulates the energy-efficient spiking neural networks (SNNs) of the human brain. This research evaluates artificial-to-spiking neural network (ANN-to-SNN) conversion as a practical methodology to perform energy-efficient heartbeat classification on the state-of-the-art Intel Loihi neuromorphic processor. In particular, a spiking 1D-convolutional neural network (1D-CNN) model is designed through ANN-to-SNN conversion with SNN-Toolbox to identify arrhythmias on Loihi. Insights into the conversion process are gained through experimentation with accuracy-latency tradeoffs, neuron reset mechanisms, and weight and bias values. These insights enable the spiking 1D-CNN to be optimized for low latency and high accuracy on Loihi, and then compared to an architecturally identical artificial neural network (ANN) on Intel Core i7 CPU, Intel Neural Compute Stick 2, and Google Coral Edge TPU devices in terms of accuracy, latency, and energy performance. Across five classes, the spiking 1D-CNN is found to reach an accuracy and macro-averaged F1 score of 97.8% and 87.9%, respectively, compared to 98.4% and 90.8% for the ANN. Additionally, with the lowest dynamic power across devices, Loihi is estimated to provide a 28 times lower energy-delay product for the model versus the CPU baseline. However, with the highest latency across devices, Loihi is also estimated to result in a 1.5 times and 110 times higher energy-delay product versus the Intel Neural Compute Stick 2 and Google Coral Edge TPU, respectively. This higher latency is determined to result from x86 core-to-host I/O and x86 core-based management bottlenecks. From these findings, insights are provided regarding future directions for practical neuromorphic computing.


Share

Citation/Export:
Social Networking:
Share |

Details

Item Type: University of Pittsburgh ETD
Status: Unpublished
Creators/Authors:
CreatorsEmailPitt UsernameORCID
Buettner, Kylekrb115@pitt.edukrb115
ETD Committee:
TitleMemberEmail AddressPitt UsernameORCID
Committee ChairGeorge, Alanalan.george@pitt.edu
Committee MemberKubendran, Rajkumarrajkumar.ece@pitt.edu
Committee MemberAkcakaya, Muratakcakaya@pitt.edu
Date: 13 June 2021
Date Type: Publication
Defense Date: 29 March 2021
Approval Date: 13 June 2021
Submission Date: 19 March 2021
Access Restriction: 1 year -- Restrict access to University of Pittsburgh for a period of 1 year.
Number of Pages: 52
Institution: University of Pittsburgh
Schools and Programs: Swanson School of Engineering > Electrical and Computer Engineering
Degree: MS - Master of Science
Thesis Type: Master's Thesis
Refereed: Yes
Uncontrolled Keywords: Heartbeat Classification, Neural Network Hardware, Neuromorphic Computing, Performance Analysis, Spiking Neural Networks
Date Deposited: 13 Jun 2021 18:28
Last Modified: 13 Jun 2021 18:28
URI: http://d-scholarship.pitt.edu/id/eprint/40401

Metrics

Monthly Views for the past 3 years

Plum Analytics


Actions (login required)

View Item View Item