1 of 12

Adaptive Edge Offloading for Image Classification�Under Rate Limit

Jiaming Qiu

qiujiaming@wustl.edu

​

Joint work with Ruiqi Wang, Ayan Chakrabarti, Roch Guérin, Chenyang Lu

1

2 of 12

System Setup

2

Camera

Edge Server

 

 

Token

Bucket

Network

Conveyor Belt

  • Cameras taking and classifying images of objects for subsequent actuation
    • Objects coming from a conveyor belt with complex arrival patterns
    • Embedded “weak” classifiers perform local classification
  • Edge Server with “strong” classifier as shared backup for low confidence local results
    • Network resources shared by multiple cameras
    • Token bucket as rate control mechanism to regulate network load

3 of 12

Problem Formulation

  •  

3

 

 

4 of 12

Key Challenges

  • Estimating immediate reward
    • Images have different offloading reward
    • Reward depends on confidence of both classifiers, but must be estimated using only the weak classifier output
  • Predicting future reward
    • Future reward depends on classifier confidence and arrival pattern of future objects
  • Offloading policy complexity
    • Needs to be able to run on local compute resources and complete within a short amount of time (before next object arrival)

4

Future Objects

Future Reward

Low

High

Objects Inter-Arrival Time

2s

3s

1s

1s

5 of 12

Solution

  •  

5

6 of 12

Evaluation

Correlation in Classification Confidence

6

  • DQN learns about output correlation and adapts its offloading policy

Offloading Metric

 

 

 

7 of 12

Evaluation

Correlation in Classification Confidence

7

Average Top-5 Loss

 

 

 

 

 

  • DQN consistently outperforms MDP and Baseline

8 of 12

Evaluation

Correlation in Arrival Rate

8

  • DQN knows to expect rate changes and adapts its offloading policy

Offloading Metric

 

 

9 of 12

Evaluation

Correlation in Arrival Rate

9

  • DQN again consistently outperforms MDP and Baseline

 

Average Top-5 Loss

 

 

10 of 12

Evaluation

Runtime Efficiency

10

  • DQN incurs negligible overhead in the real-time processing pipeline

11 of 12

Summary

  • A lightweight DQN model (5 ×64 multi-layer perceptron) is capable of learning:
    • A complex image arrival process
    • A complex classifier output process
    • The impact of the state of a token bucket and its constraints

to make offloading decisions that optimize classification accuracy

11

12 of 12

Thank you!

Questions?

12