Generic Neural Elastic Search

Han Xiao

• Working at Tencent AI
• Successful OpenSource projects he worked on
  • Fashion-MNIST
  • Bert-as-service

Tencent != 10¢

• From Internet to Wechat
• Basically Google for China + they managed to be good at social media and messaging
• They have a lot of data and they can probably use it, because there is no GDPR in China

Tencent AI

• 70 AI scientists + 300 app developers (== ML Engineers)
• The #1 in AI publications in China, yet still 1/10 of Google papers

Source: github/GNES

GNES ‒ Generic Neural Elastic Search

• One framework to train text, image and video search

• Should be used as TF, just everything deployed in a different place
• A bit like KuberFlow/AirFlow, but optimized for Search scenario

Premises

• Cloud native
• Semantic search (using DNN)
• End2end

Obstacles

• What is the distance metric for doc vectors?
• How to handle the difference between short / long texts (video, images)

The Idea

• Minimum information unit × minimum semantic unit × optimum semantic unit
• a word × a sentence × ???
• a pixel × a 64×64 patch × ???
• a pixel in 1 frame × a 2-3 sec shot × ???
• What are the optima?
  • Do an experiment!
    • => determined by ONE preprocessor
• Models change too quickly
  • => plug them in as a Docker container

Specs

• Everything is its own microservice
• All of the components are defined by a YAML file => immutable code, just change the YAML

Overview of possible parts
Source: github/GNES

Questions

• Are you happy with the “All as microservice” design?
• Do you have any idea, what is the overhead?
  • It adds overhead if the processing of each component is small
    • => great for images or better video (just FFMPeg)
• Is the presentation somewhere?
  • Maybe
• Can it be put into lambda functions?
  • Not for now
  • Not because of the loading

---

• How to address multiple languages – when preprocessing is completely different?
  • It is your job to write a specific YAML file that preprocesses it
• How to handle huge models in Docker images
  • Just put everything into the container
• What is the view in China on big OpenSource contributions?
  • They do not get much support for OpenSource (a lot of papering, what to publish etc.)
  • 30 % work the relationship making
• Challenges for Europe with regards to AI
  • A lot of Chinese AI is focused on the consumer (you can test something on this milion of data and that milion of data)
  • => not much 0 to 1 research, a lot of 1 to N research
  • Don’t try to copy B2C of US and China
  • Try to focus on the middle man, companies, which can change the world a lot, but are not “shiny” (no coverage)
    • Upgrade the traditional industry! Keep Europe great still!