Min, Chuhan
(2020)
HMC-Based Accelerator Design For Compressed Deep Neural Networks.
Doctoral Dissertation, University of Pittsburgh.
(Unpublished)
Abstract
Deep Neural Networks (DNNs) offer remarkable performance of classifications and regressions in many high dimensional problems and have been widely utilized in real-word cognitive applications. In DNN applications, high computational cost of DNNs greatly hinder their deployment in resource-constrained applications, real-time systems and edge computing platforms. Moreover, energy consumption and performance cost of moving data between memory hierarchy and computational units are higher than that of the computation itself. To overcome the memory bottleneck, data locality and temporal data reuse are improved in accelerator design. In an attempt to further improve data locality, memory manufacturers have invented 3D-stacked memory where multiple layers of memory arrays are stacked on top of each other. Inherited from the concept of Process-In-Memory (PIM), some 3D-stacked memory architectures also include a logic layer that can integrate general-purpose computational logic directly within main memory to take advantages of high internal bandwidth during computation.
In this dissertation, we are going to investigate hardware/software co-design for neural network accelerator. Specifically, we introduce a two-phase filter pruning framework for model compression and an accelerator tailored for efficient DNN execution on HMC, which can dynamically offload the primitives and functions to PIM logic layer through a latency-aware scheduling controller.
In our compression framework, we formulate filter pruning process as an optimization problem and propose a filter selection criterion measured by conditional entropy. The key idea of our proposed approach is to establish a quantitative connection between filters and model accuracy. We define the connection as conditional entropy over filters in a convolutional layer, i.e., distribution of entropy conditioned on network loss. Based on the definition, different pruning efficiencies of global and layer-wise pruning strategies are compared, and two-phase pruning method is proposed. The proposed pruning method can achieve a reduction of 88% filters and 46% inference time reduction on VGG16 within 2% accuracy degradation.
In this dissertation, we are going to investigate hardware/software co-design for neural network accelerator. Specifically, we introduce a two-phase filter pruning framework for model compres- sion and an accelerator tailored for efficient DNN execution on HMC, which can dynamically offload the primitives and functions to PIM logic layer through a latency-aware scheduling con- troller.
In our compression framework, we formulate filter pruning process as an optimization problem and propose a filter selection criterion measured by conditional entropy. The key idea of our proposed approach is to establish a quantitative connection between filters and model accuracy. We define the connection as conditional entropy over filters in a convolutional layer, i.e., distribution of entropy conditioned on network loss. Based on the definition, different pruning efficiencies of global and layer-wise pruning strategies are compared, and two-phase pruning method is proposed. The proposed pruning method can achieve a reduction of 88% filters and 46% inference time reduction on VGG16 within 2% accuracy degradation.
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: |
20 November 2019 |
| Approval Date: |
29 January 2020 |
| Submission Date: |
21 November 2019 |
| Access Restriction: |
No restriction; Release the ETD for access worldwide immediately. |
| Number of Pages: |
99 |
| 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: |
Neural network, HMC, model compression, software/hardware co-design |
| Date Deposited: |
29 Jan 2020 16:13 |
| Last Modified: |
29 Jan 2020 16:13 |
| URI: |
http://d-scholarship.pitt.edu/id/eprint/37869 |
Metrics
Monthly Views for the past 3 years
Plum Analytics
Actions (login required)
 |
View Item |
![[feed]](/images/feed-icon-32x32.png){kind=icon}