Title

Evolutionary Search Algorithm for the Optimized Dataflow Engine in Deep Learning

Alternative Title

심층신경망 연산에 최적화된 NPU 구조를 위한 진화 알고리즘

DGIST Authors

Jooyeon Lee ; Jaeha Kung ; Sungjin Lee

Advisor

궁재하

Co-Advisor(s)

Sungjin Lee

Issued Date

2021

Awarded Date

2021/02

Citation

Jooyeon Lee. (2021). Evolutionary Search Algorithm for the Optimized Dataflow Engine in Deep Learning. doi: 10.22677/thesis.200000366431

Type

Thesis

Subject

Dataflow, Evolutionary algorithm, Simulator, 심층 신경망, 진화 알고리즘, 최적화, 데이터 흐름, 성능 시뮬레이션

Table Of Contents

Ⅰ. Introduction 1
Ⅱ. Background 1
2.1 Multilayer Neural Networks 1
2.2 Convolution Neural Network (CNN) 3
2.3 Detailed Description of Each Layer in Convolution Neural Network 3
2.3.1 Convolution Layer 3
2.3.2 Depthwise Separable Convolution Layer 4
2.3.3 Non-Linearity 5
2.3.3 Pooling Layer 5
Ⅲ. Prior work: Dataflow Simulation Engine 6
3.1 Definition of Dataflow 6
3.2 NPU Dataflow Simulators 6
3.3 Nested Loops in Computing DNN Layers 8
3.4 Semantic Representation 9
3.5 Examples of Previous Dataflow Models 11
3.5.1 Weight Stationary Dataflow 11
3.5.2 Row Stationary Dataflow 12
3.6 Open-source Dataflow Simulator 13
Ⅳ. Finding Optimized Dataflow with Evolutionary Algorithm 13
4.1 Evolutionary Algorithm for Searching the Optimized Dataflow 13
4.2 Details on Mutation Operators 15
Ⅴ. Experimental Results 16
Ⅵ. Conclusion 19

URI

http://dgist.dcollection.net/common/orgView/200000366431
http://hdl.handle.net/20.500.11750/16728

DOI

10.22677/thesis.200000366431

Degree

Master

Department

Information and Communication Engineering

Publisher

DGIST

Show Full Item Record