In this talk, I describe our journey of developing a surface inspection AOI machine and discuss how a seemingly simple problem can be deceptively hard and end up taking longer than anticipated.
UiPath Community: AI for UiPath Automation Developers
A Simple AI Problem That Wasn’t: A Smart Manufacturing Case Study
1. A Simple AI Problem
That Wasn’t
Wei-Chao Chen 陳維超
Chief Digital Officer & SVP, Inventec Inc.
Co-Founder, Skywatch Inc.
chen.wei-chao@inventec.com
For ICPAI 2020, December 2020
5. 5
Inventec Confidential
o S1: Write a fancy diff program
Find Visual Defects on Laptops
How hard can it be?
Golden Defective
6. 6
Inventec Confidential
o S1: Write a fancy diff program
o Variance between capture
Find Visual Defects on Laptops
How hard can it be?
7. 7
Inventec Confidential
o S1: Write a fancy diff program
o Variance between capture
o S2: Use an object detector
Find Visual Defects on Laptops
How hard can it be?
9. 9
Inventec Confidential
o S1: Write a fancy diff program
o Variance between capture
o S2: Use an object detector
o Lots of data to label
Find Visual Defects on Laptops
How hard can it be?
10. 10
Inventec Confidential
o S1: Write a fancy diff program
o Variance between capture
o S2: Use an object detector
o Lots of data to label
o S3: Let us capture and label the data anyways
Find Visual Defects on Laptops
How hard can it be?
11. 11
Inventec Confidential
o S1: Write a fancy diff program
o Variance between capture
o S2: Use an object detector
o Lots of data to label
o S3: Let us capture and label the data anyways
o You forgot to look at the sides
Find Visual Defects on Laptops
How hard can it be?
12. 12
Inventec Confidential
o S1: Write a fancy diff program
o Variance between capture
o S2: Use an object detector
o Lots of data to label
o S3: Let us capture and label the data anyways
o You forgot to look at the sides
o S4: Build a proper machine
Find Visual Defects on Laptops
How hard can it be?
13. 13
Inventec Confidential
S4: Build a proper machine
Surely it took a while and a bit of fortune
Inventec Laptop AOI Machine, v2.0, GTC 2020
14. 14
Inventec Confidential
o S1: Write a fancy diff program
o Variance between capture
o S2: Use an object detector
o Lots of data to label
o S3: Let us capture and label the data anyways
o You forgot to look at the sides
o S4: Build a proper machine
o Well, but the product is obsolete before the model is ready
Find Visual Defects on Laptops
How hard can it be?
>6 Months!!
99+% yield, 1000s daily volume
15. 15
Inventec Confidential
o S1: Write a fancy diff program
o Variance between capture
o S2: Use an object detector
o Lots of data to label
o S3: Let us capture and label the data anyways
o You forgot to look at the sides
o S4: Build a proper machine
o Well, but the product is obsolete before the model is ready
o S5: Use less labels with semi-supervised algorithm
Find Visual Defects on Laptops
How hard can it be?
17. 17
Inventec Confidential
o S1: Write a fancy diff program
o Variance between capture
o S2: Use an object detector
o Lots of data to label
o S3: Let us capture and label the data anyways
o You forgot to look at the sides
o S4: Build a proper machine
o Well, but the product is obsolete before the model is ready
o S5: Use less labels with semi-supervised algorithm
o Fine, but the accuracy is worse than expected
Find Visual Defects on Laptops
How hard can it be?
18. 18
Inventec Confidential
o S1: Write a fancy diff program
o Variance between capture
o S2: Use an object detector
o Lots of data to label
o S3: Let us capture and label the data anyways
o You forgot to look at the sides
o S4: Build a proper machine
o Well, but the product is obsolete before the model is ready
o S5: Use less labels with semi-supervised algorithm
o Fine, but the accuracy is worse than expected
o S6: Wait, let us check if humans are better at it
o Well, turned out they are not
Find Visual Defects on Laptops
How hard can it be?
20. 20
Inventec Confidential
o Passing criteria can vary across inspectors
o Label quality, acceptance criteria hard to define
The Weakest Link
To err is human
Visual
Criteria
Time
Product
Type
21. 21
Inventec Confidential
o S1: Write a fancy diff program
o Variance between capture
o S2: Use an object detector
o Lots of data to label
o S3: Let us capture and label the data anyways
o You forgot to look at the sides
o S4: Build a proper machine
o Well, but the product is obsolete before the model is ready
o S5: Use less labels with semi-supervised algorithm
o Fine, but the accuracy is worse than expected
o S6: Wait, let us check if humans are better at it
o Well, turned out they are not
o S7: Ok we beat the human, hurray!
Find Visual Defects on Laptops
How hard can it be?
22. 22
Inventec Confidential
o S1: Write a fancy diff program
o Variance between capture
o S2: Use an object detector
o Lots of data to label
o S3: Let us capture and label the data anyways
o You forgot to look at the sides
o S4: Build a proper machine
o Well, but the product is obsolete before the model is ready
o S5: Use less labels with semi-supervised algorithm
o Fine, but the accuracy is worse than expected
o S6: Wait, let us check if humans are better at it
o Well, turned out they are not
o S7: Ok we beat the human, hurray!
o Your machine is too expensive
Find Visual Defects on Laptops
How hard can it be?
23. 23
Inventec Confidential
o Talks
o “Edge AI Smart Manufacturing - Defect Detection and Beyond”, T.
Chen, W-C. Chen, in NVIDIA GTC 2019
o “Toward Taming the Training Data Complexity in Smart
Manufacturing”, D. Tan, H-H. Lee, Y-C. Chen, W-C. Chen, T. Chen,
in NVIDIA GTC 2020
o Papers
o “TrustMAE: A Noise-Resilient Defect Classification Framework
using Memory-Augmented Auto-Encoders with Trust Regions”, D.
Tan, Y-C. Chen, T. Chen, W-C. Chen, in WACV 2021
o “Demystifying Data and AI for Manufacturing: Case Studies from a
Major Computer Maker”, Y-C. Chen et al., in APSIPA Trans 2021.
References
Contact: chen.wei-chao@inventec.com
24. A Simple AI Problem
That Wasn’t
Wei-Chao Chen 陳維超
Chief Digital Officer & SVP, Inventec Inc.
Co-Founder, Skywatch Inc.
chen.wei-chao@inventec.com
For ICPAI 2020, December 2020