Roffe, Seth
(2023)
On the Reliability of Neuromorphic, Event-Based Systems for Space.
Doctoral Dissertation, University of Pittsburgh.
(Unpublished)
Abstract
Neuromorphic, event-driven systems can be separated into two main sections: neuromorphic vision and neuromorphic processing. Both are remarkably efficient methods that aim to offer a new archetype of computing. The shared concept between the two is to process or sense in the temporal domain. Event-based vision sensors replicate biological retinas to make use of their high power efficiency, sparse output representation, and large dynamic range. Similarly, neuromorphic processors are modelled after the human brain to simulate how neurons fire and learn. This computational model improves power efficiency, enables native machine-learning capabilities, and overcomes the von Neumann memory bottleneck.
This research designs, creates, and evaluates a full system for reliable sensor processing within a neuromorphic classification system from end to end. This evaluation involves ensuring that the failure modes and reliability of a neuromorphic system are known at every step from sensor data, to processing data, to output data. The matrix-multiplication kernel was chosen as a common algorithm needed for ML/CV applications and evaluated for its reliability and efficiency under different dependable-computing techniques. Given the results from this evaluation, a neuromorphic vision sensor was chosen for further study due to its promise in low-power ML/CV capabilities and low data rate. This research provides the first radiation test data to observe and model the effects induced by radiation. The Event-Based Radiation-Induced-Noise Simulation Environment (Event-RINSE) is proposed as a fault injector to simulate the modeled neutron effects on event data without the need for radiation testing. Finally, a neuromorphic classification method, the Hierarchy of Event-Based Time-Surfaces (HOTS) is studied for use in a radiative environment such as space to build off of the previous two experiments. Specifically, how the Time Surface features and other common neuromorphic computations such as time delays respond to radiation noise, and how upsets affect classification accuracy, are evaluated. Given these results, methods to create a more reliable neuromorphic architecture for use in hazardous environments are proposed. Each section provides a piece of a complete neuromorphic classification system. This research provides a starting point to realizing a reliable, fully neuromorphic sensing and processing system for future spacecraft.
Share
| Citation/Export: |
|
| Social Networking: |
|
Details
| Item Type: |
University of Pittsburgh ETD
|
| Status: |
Unpublished |
| Creators/Authors: |
|
| ETD Committee: |
|
| Date: |
19 January 2023 |
| Date Type: |
Publication |
| Defense Date: |
19 October 2022 |
| Approval Date: |
19 January 2023 |
| Submission Date: |
24 October 2022 |
| Access Restriction: |
No restriction; Release the ETD for access worldwide immediately. |
| Number of Pages: |
118 |
| Institution: |
University of Pittsburgh |
| Schools and Programs: |
Swanson School of Engineering > Electrical and Computer Engineering |
| Degree: |
PhD - Doctor of Philosophy |
| Thesis Type: |
Doctoral Dissertation |
| Refereed: |
Yes |
| Uncontrolled Keywords: |
Neuromorphic Computing, Reliability, Dependability |
| Date Deposited: |
19 Jan 2023 19:12 |
| Last Modified: |
19 Jan 2023 19:12 |
| URI: |
http://d-scholarship.pitt.edu/id/eprint/43757 |
Metrics
Monthly Views for the past 3 years
Plum Analytics
Actions (login required)
 |
View Item |
![[feed]](/images/feed-icon-32x32.png){kind=icon}