cDeepArch: A Compact Deep Neural Network Architecture for Mobile - - PowerPoint PPT Presentation

▶
cdeeparch a compact deep neural network architecture for
SMART_READER_LITE
LIVE PREVIEW

cDeepArch: A Compact Deep Neural Network Architecture for Mobile - - PowerPoint PPT Presentation

Sep 16, 2022 197 likes •439 views

cDeepArch: A Compact Deep Neural Network Architecture for Mobile Sensing Kang Yang 1 , Xiaoqing Gong 1 , Yang Liu 2 , Zhenjiang Li 2 , Tianzhang Xing 1 , Xiaojiang Chen 1 , Dingyi Fang 1 1 Northwest University, China 2 City University of Hong Kong

slide-1
SLIDE 1

cDeepArch: A Compact Deep Neural Network Architecture for Mobile Sensing

Kang Yang1, Xiaoqing Gong1, Yang Liu2, Zhenjiang Li2, Tianzhang Xing1, Xiaojiang Chen1, Dingyi Fang1

1Northwest University, China 2City University of Hong Kong

1

slide-2
SLIDE 2

Camera Gyro. Acc.

…

+

Learning Technology

Motivation

slide-3
SLIDE 3

Application

?

Cognitive decline

slide-4
SLIDE 4

Application

pot cup close

  • pen

First-person view Recognizing

Cognitive aid system

  • pen
slide-5
SLIDE 5

Common design principle

Rich sensor data Recognized by learning

. . .

Applications

slide-6
SLIDE 6

Challenges

. . .

Large targets

slide-7
SLIDE 7

Big deep neural network

Challenges

  • Deep Learning

Too large

Resource-limited

slide-8
SLIDE 8

Shrunk model

Challenges

  • Deep Learning

Original model No quantitative measure on available resource conditions inaccurate

slide-9
SLIDE 9

Server

  • Long and uncontrollable latency
  • High Service cost
  • Potential privacy leakage

0101…

Any countermeasure?

slide-10
SLIDE 10

Large targets

Our solution

Context (office)

. . .

slide-11
SLIDE 11

Our solution

Context recognition Context-oriented target recognition Object recognition

+

adequate storage

large and deep network compact network compact network (Office) (computer, mouse…)

computation resource

slide-12
SLIDE 12

Our solution

Context recognition Context-oriented target recognition

+

Available resource conditions

computation energy

  • not based on designer’s experience
  • Formulation facilitated configuration
slide-13
SLIDE 13

Convolutional Neural Network

Image data Conv1 Pool1 Conv2 Pool2 FC1

  • Convolutional layer (dominant)
  • Full connected layer
  • Pooling layer
slide-14
SLIDE 14

(W+2P)*(W+2P)

C W

Wo *Wo F P S

Formulation facilitated configuration

Selected

slide-15
SLIDE 15

computation

From computation to resource cost

Conv1:16 Conv2:32 fc:5 Conv1:64 Conv2:128 fc:5

a small scale network

: computation : actual resource consumption

designed network resource(energy)

slide-16
SLIDE 16

Now…

Context recognition Context-oriented target recognition

+

  • Formulation facilitated configuration
  • From formulation to estimate the resource consumption

Object recognition

  • Recognition task decomposition
slide-17
SLIDE 17

! ≤ #$% − ⁄ 1 ) ⁄ ) 2

Enhancement: Convolutional layer

Original model

Conv1 Conv2 Conv3

Separated model

Conv2 Conv3 Conv1a Conv1b

slide-18
SLIDE 18
slide-19
SLIDE 19
slide-20
SLIDE 20

Evaluation

slide-21
SLIDE 21
  • Dataset:
  • Context recognition:

§ MIT Place2 (related to the daily contexts)

  • Object recognition:

§ Cifar10 § Cifar100 (20 classes associated contexts)

Experiments setup

slide-22
SLIDE 22
  • Overall performance

Evaluation results

10 targets 20 targets

slide-23
SLIDE 23

Conclusion 1, 2, 3

  • 1. Large targets Decompose recognition task
  • 2. Systematic way to configure network Execution
  • verhead formulation facilitated configuration
  • 3. Enhancement techniques