Wang, Yandan
(2020)
Efficient and Robust Neuromorphic Computing Design.
Doctoral Dissertation, University of Pittsburgh.
(Unpublished)
Abstract
In recent years, brain inspired neuromorphic computing system (NCS) has been intensively studied in both circuit level and architecture level. NCS has demonstrated remarkable advantages for its high-energy efficiency, extremely compact space occupation and parallel data processing. However, due to the limited hardware resources, severe IR-Drop and process variation problems for synapse crossbar, and limited synapse device resolution, it’s still a great challenge for hardware
NCS design to catch up with the fast development of software deep neural networks (DNNs). This dissertation explores model compression and acceleration methods for deep neural networks to save both memory and computation resources for the hardware implementation of DNNs. Firstly, DNNs’ weights quantization work is presented to use three orthogonal methods to learn synapses with one-level precision, namely, distribution-aware quantization, quantization regularization and bias tuning, to make image classification accuracy comparable to the state-ofthe-art. And then a two-step framework named group scissor, including rank clipping and group connection deletion methods, is presented to address the problems on large synapse crossbar
consuming and high routing congestion between crossbars.
Results show that after applying weights quantization methods, accuracy drop can be well controlled within negligible level for MNIST and CIFAR-10 dataset, compared to an ideal system without quantization. And for the group scissor framework method, crossbar area and routing area could be reduced to 8% (at most) of original size, indicating that the hardware implementation area has been saved a lot. Furthermore, the system scalability has been improved significantly.
Share
| Citation/Export: |
|
| Social Networking: |
|
Details
| Item Type: |
University of Pittsburgh ETD
|
| Status: |
Unpublished |
| Creators/Authors: |
|
| ETD Committee: |
|
| Date: |
29 January 2020 |
| Date Type: |
Publication |
| Defense Date: |
13 November 2019 |
| Approval Date: |
29 January 2020 |
| Submission Date: |
19 November 2019 |
| Access Restriction: |
No restriction; Release the ETD for access worldwide immediately. |
| Number of Pages: |
97 |
| 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; large neural networks; low-rank approximation; memristor; random number generator; stochastic switching; weight quantization, deep neural networks |
| Date Deposited: |
29 Jan 2020 16:38 |
| Last Modified: |
29 Jan 2020 16:38 |
| URI: |
http://d-scholarship.pitt.edu/id/eprint/37797 |
Metrics
Monthly Views for the past 3 years
Plum Analytics
Actions (login required)
 |
View Item |
![[feed]](/images/feed-icon-32x32.png){kind=icon}