Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[20220515] Weekly AI ArXiv 만담 - 51회차 #51

Closed
jungwoo-ha opened this issue May 15, 2022 · 4 comments
Closed

[20220515] Weekly AI ArXiv 만담 - 51회차 #51

jungwoo-ha opened this issue May 15, 2022 · 4 comments

Comments

@jungwoo-ha
Copy link
Owner

jungwoo-ha commented May 15, 2022

News

ArXiv

  • Unifying Language Learning Paradigms

    • LM을 위한 통합 학습기법: Dec only도 Enc-Dec 도 모두 가능 (from Google Research, PaLM팀)
    • Pretrainig task: Mixture of Denoisers (S, R, X)
    • Denosing mode 변환을 위해 paradigm token 활용
    • 그래서 Finetuning 에 서도 in-context learning에서도 더 좋은 성능
    • 아쉬운 점: 얘는 비교를 하려면 Chinchilla 비슷한 크기로 맞춰서 비교했어야..
    • 적당한 크기 (수십B 이하) 에서 파라미터를 굉장히 잘활용해서 좋은 LM을 만드는 기법들 (좋은 소식)
      image
      image
  • A Generalist Agent

    • Gato: 604개의 RL게임, Vision 기반 Robot control, Text, image-text 캡셔닝 등 604개의 task를 동일한 웨이트 하나의 1.2B Transformer 로 다루는 제목 그대로 generalist agent (from Deepmind)
    • 모든 입력은 token으로 변환 (텍스트, 이미지, discrete value, continuous value), Voc size는 대략 35000
    • output은 target token prediction 그래서 image 와 agent observation token prediction task는 없음
    • 이미지는 ResNet ecoder로 임베딩, 나머지는 모두 lookup table 스타일 임베딩
    • task 구분없이 동일한 형태의 token 시퀀스로 다 집어넣고 학습
    • 동일 도메인내 다양한 태스크 구분을 위해서는 prompt 를 활용
    • 16 x 16 TPU v3 512batch 1M step 4일간 학습 (의외로 computation 많이 안들어가는 듯)
    • 나머지 각 도메인 태스크 결과는 논문에서 확인 가능하십니당.
      image
      image
      image
  • 눈여겨볼 논문들

@ghlee0304
Copy link

ghlee0304 commented May 15, 2022

Arxiv

@kimyoungdo0122
Copy link

kimyoungdo0122 commented May 15, 2022

News

  • Hugging Face가 1억 달러 시리즈 C 투자를 유치했습니다

    • Hugging Face reaches $2 billion valuation to build the GitHub of machine learning
    • 시리즈 A에서는 1천 500만 $, B에서 4천만 $, C에서 1억 $ 투자를 받았고 현재는 기업가치가 20억 달러에 이른다는 내용
    • 허깅페이스의 BM은 SaaS?
    • 수많은 pre-trained model을 보유한 허깅페이스가 AI community에 이미 엄청난 영향력을 갖고 있는데,
    • infrastructure solution이 곁들여진다면 Large-scale auto ml 분야와 같은 enterprise에도 큰 영향력을 가질 것 같습니다
  • Apple의 Machine Learning Chief인 Ian Goodfellow가 퇴사했다고 합니다

    • Apple’s chief of machine learning quits over return-to-office policy
    • 애플이 코로나 상황이 끝나간다고 느꼈는지, 5월 말에는 주 3회 오피스 의무 출근을 추진하고 있다고 합니다.
    • 그리고 점차 풀 오피스 출근으로 전환하는 것에 이안 굿펠로우는 물론 여러 애플 임직원들이 반발하고 있다고 하네요
    • 포스트 코로나가 다가오면서 엔지니어 직군을 포함해 많은 사람들이 일하는 방식에 대해서 갈등이 있을 듯합니다

@veritas9872
Copy link

Better plain ViT baselines for ImageNet-1k

Arxiv: https://arxiv.org/pdf/2205.01580.pdf
GitHub: https://github.com/google-research/big_vision/tree/main/big_vision

ViT 학습이 기존 통념보다 잘 된다고 합니다. 기존 CNN에 사용하던 테크닉 없이 단순한 학습 방법을 적용했을 때 오히려 ViT 모델이 성능이 더 좋다는 구글의 연구 결과입니다. ViT 모델이 문제없이 ImageNet top-1 80%를 300 epoch 만에 찍는 것을 보여줍니다. 결론은 너무 어렵게 생각하지 말자는 것...

Screenshot (52)

Screenshot (53)

CoCa: Contrastive Captioners are Image-Text Foundation Models

Screenshot (54)

ImageNet 새로운 SOTA가 등장했습니다. 구글에서 Image to Text 모델을 학습한 결과 모든 라벨을 text 형태로 촐력하게 하여 zero-shot accuracy 86.3%와 fine-tuning 결과 91.0%라는 SOTA를 세웠습니다. 결론은 역시 돈 많은 자가 승리한다는 것...

Screenshot (56)

이번 주는 너무 새로 나온 내용이 많아서 Yannic Kilcher의 YouTube 공유해드립니다.

ML News Part 1: https://www.youtube.com/watch?v=pwSnC8jlh50
ML News Part 2: https://www.youtube.com/watch?v=smUHQndcmOY

@veritas9872
Copy link

veritas9872 commented May 15, 2022

Tensorboard / W&B 대신 사용할 수 있는 AimStack 공유해드립니다. https://aimstack.io/
W&B와는 다르게 오픈소스입니다!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

4 participants