Edge AI Design Strategy for C2C E-Commerce Applications

Explore edge AI design strategies for client-to-client e-commerce applications in this 35-minute talk by Yuji Oshima from the Linux Foundation. Delve into the implementation of client-side machine learning in Mercari, examining its advantages and development process. Discover four key production considerations, compare server-side and client-side approaches, and analyze the Mercari Lens model architecture. Learn about inference on streaming media, pre/post-processing techniques, and the use of MediaPipe. Investigate process pipeline definition, inference optimization, and validation methods. Discuss the challenges of web-based tools and explore architecture for native applications, including the potential for web migration with native-like performance. Gain valuable insights into cutting-edge AI implementation for e-commerce platforms.

Intro
Agenda
Background replacement on Google Meet
Client-side ML in Mercari
Why Client-side ML
How to develop Client-side ML
For more advanced use cases
Four key points concerning production
Is the Server-side enough?
In the case of Mercari Lens
Model architecture
Inference on Streaming Media
Pre/Post Process
Introduce MediaPipe
Define a process pipeline
Inference with the pipeline
Optimize the pipeline
Validation
Tools for web
Unfortunately, NO
Architecture for native application
We could migrate to web!
Native like performance on web
Conclusion