AIGS AI Seminar : 오태현 교수님( from POSTECH )

 

Seminars

학교에서는 종종 AI 세미나가 열리고, 그 중에는 AI 대학원 에서 진행하는 세미나 수업으로 인해 열리는 경우가 대부분이다.
그냥 흘려 보내기에는 아까우니까 정리를 하면 좋겠다는 생각이 들었다.

Seminar Info

  • Speaker : POSTECH 오태현 교수님
  • Topic : Data efficient ML/DL/AI

Memos

Intro

  • Big model size –> reason of success became the bottleneck.
  • 175B parameters –> 1TB model size –> 12M$ per training!
  • But still not enough. because …
    • 175B GPT3 $«$ 125T brain synapses.
    • HW advance saturated
    • Always lacking labeled dataset
    • Over parametrized, under determined.
  • 1.5 hr for a single img annotation, but very poor quality even costly amount of money.
  • sometimes very complex annotation logic needed,
  • sometimes there is no way to annotate what
  • Toolkits
    • synthetic dataset
    • prior knowledge
    • semi-supervised methods
    • heterogeneous dataset
    • applying classical methods
    • etc

Case 1

Problem situation

  • no proper dataset for magnified video images
  • Fourier transform cannot model new signals appearing, like occlusion.

Solution : make synthetic data by immitating movements.

  • crop background/objects from COCO/PASCAL VOC
  • apply simple translation, immitating movements
  • why? no context needed.

Case 2

  • Entire model param space is inf
  • add inductive bias (prior information) to narrow range!
  • blackbox –> constraint –> blackbox format
  • constraints? physics law, etc …

Case 3

Neural inverse knitting : 자수를 보고 만드는 법을 역으로 생성.

  • 2000 knitting patterns takes 2 monthes. too long.
  • real data를 synthetic data 처럼 만들어서 regularize!
  • 보통은 syn2real을 시도하는데, 반대로 함.
  • real data : distribution very random / syn data : distribution pure
  • real2syn –> neutralize variance

Case 4

Fedpara 논문.
$rank(W)$ 가 parameter capacity와 연결된다고 생각함.
rank 가 $R_1$, $R_2$ 인 두 matrices 로 나눠서 Hadamard product.
이렇게 하면 $rank(W) = R_1 R_2 $ 가 되므로 quadratic 하게 늘어남!

Conclusion

  • synthetic data, syn2real 은 충분히 적용 할만하다.
  • 멋있다

wbjeon2k

Pursuite for Progress.

This work is licensed under a Attribution-NonCommercial 4.0 International license. Attribution-NonCommercial 4.0 International