4. About me
• Profile
• Kosuke Kuzuoka
• 23 years old
• Experience
• June 2018 – Present
AI Research Engineer
DeNA Co., Ltd.
• March 2017 – June 2018
R&D Manager
Photoruction, inc.
• Interests
• Self Driving Cars
• Computer Vision
• Cars, especially Tesla
4
5. 5
Brief Intro to Object Detection
• An active research area among
computer vision community
• Task is detecting objects
(like cats) in an image
• Modern algorithms heavily
rely on deep learning
• Modern algorithms have
million of parameters
Photo by Paul Hanaoka on Unsplash
6. Brief Intro to Object Detection
6
• Detector takes an image, find object-ish
regions and determines what the object is
• Detector has hyper-parameters,
that need to be carefully tuned
• Detector learns through propagating
losses backwards
• Learning takes hours on modern GPUs
Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
https://arxiv.org/abs/1506.01497
7. 7
A cat is detected as a cat,
hence it’s a true positive.
Wrongly detected as cats,
hence they are false positives
Photo by Paul Hanaoka on Unsplash
9. AI Development Pipeline
9
1. Train, validate and test AI model on GPU machine
2. Check the mAP (evaluation protocol), visualize results etc..
3. Adjust hyper-param then go back to 1.
10. AI Development Pipeline
10
1. Train, validate and test AI model on GPU machine
2. Check the mAP (evaluation protocol), visualize results etc..
3. Adjust hyper-param then go back to 1.
11. Problems
11
• Error-prone process (misspelling commands, etc.)
• Going back and forth between EC2 instances…
• Inefficient process, like drawing boxes, uploading
to third party app for visualization etc.
12. Solutions
12
• Work harder and harder...
• Automation via workflow engine
• Internal tool to evaluate, visualize
and analyze models
13. Solutions
13
• Work harder and harder...
• Automation via workflow engine
• Internal tool to evaluate, visualize
and analyze models
15. Why Serverless?
15
• Easy to deploy and maintain
• You can think about servers, less ☺
• New features can be easily added
• Collaborations made easy
• You can use python
Image source: https://serverless.com/
16. Introducing Kaiseki-kun
16
• Maintainable and scalable analysis tool
• Automates boring manual processes
• Visualization through a user-friendly web viewer
• Lets you focus on AI work (parameter tuning, etc.)
20. 20
5 true positives and 1 false negative.
Adjusting the threshold further...
21. 21
The model wasn’t confident enough.
Perhaps add more data on the category?
22. Serverless is Awesome!
22
• Developed in only two weeks!
• Adding more features using spare
time (while training models, etc.)
• Task driven development enables
work on both AI and full-stack
24. Wrap up
24
• Serverless enables us develop both
AI model and full-stack
• Boring process like visualizing outputs
and evaluating models are automated
• More features on their way...
• Work smarter, not harder ☺
25. Becoming a Full-Stack AI Engineer like..
Photo by Tran Mau Tri Tam on Unsplash Photo by Stephanie LeBlanc on Unsplash