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

[20211030] Weekly AI ArXiv 만담 #28

Closed
jungwoo-ha opened this issue Oct 30, 2021 · 7 comments
Closed

[20211030] Weekly AI ArXiv 만담 #28

jungwoo-ha opened this issue Oct 30, 2021 · 7 comments

Comments

@jungwoo-ha
Copy link
Owner

jungwoo-ha commented Oct 30, 2021

@ghlee3401
Copy link

ghlee3401 commented Oct 31, 2021

@nick-jhlee
Copy link

행사 소개

  • Conference on the Mathematical Theory of Deep Neural Networks가 곧 virtual로 열려요... (11/4~11/5)
    https://deepmath-conference.com/
    registration deadline: 11/4까지!

    • Theoretical ML/DL에 관한 최신 트렌드를 알 수 있는 작은 conference!
    • SGD, Implicit Regularization, Learning dynamics... 등등
  • 한국인공지능학회ㅣ LG AI연구원 추계공동학술대회가 곧 virtual로 열려요... (11/4~11/5)
    http://aiassociation.kr/Conference/ConferenceView.asp?AC=0&CODE=CC20210801&CpPage=123#CONF
    registration deadline: 11/3까지!

    • Prof. Honglak Lee, Prof. Roger Grosse, Prof. Neil Lawrence!! 초청/기조 강연
    • KAIST, Yonsei, Chung-ang AI 대학원 교수님들 참여
    • 등등

@nick-jhlee
Copy link

nick-jhlee commented Oct 31, 2021

논문들

  • Learning in High Dimension Always Amounts to Extrapolation (arXiv 2021)

    • FMeta-AIR & NYU (feat. LeCun 아저씨)
    • "Universal approximation is the reason for success" ==> Hell no! Curse of dimensionality?? ==> Really?
    • Definition. interpolation은 이미 있는 데이터 sample들의 convex hull안에 있는 새로운 sample을 가지고 놀 때이다...
    • 근데 dimension이 크면 (당연히) x in conv(X)일 확률은 0으로 수렴한다... ==> 사실 우리가 interpolation이라고 하는건 다 extrapolation이었다...
    • 결론: " the behavior of a model within a training set’s convex hull barely impacts that model’s generalization performance since new samples lie almost surely outside of that convex hull"
    • My (and many other people's) opinions:
      • 근데... interpolation의 정의가 저렇게 linear해도 되는건가용,,,
      • nonlinear (Riemannian) data manifold? linear separability를 고려하면 convex hull 필요없? probabilistic models?
      • 궁금하신 분들은 논문이랑 LeCun옹의 twitter ㄱㄱ해보세용! (특히 논란과 별개로 논문은 잘 쓰여져 있어서 한번 읽어보시는거 추천드립니당...)
  • An Investigation into the Role of Author Demographics in ICLR Participation and Review (ICLR Openreview)

    • 저자 누군지 모르겠지만, 추정: MIT
    • 여기 summary: https://twitter.com/HannesStaerk/status/1454473672404652040?s=20
    • 재밌는 부분들:
      • author수가 많을수록 reviewer score가 더 높았다
      • US의 acceptance rate가 East Asia의 rate보다 1.5~2배 정도 더 높았다
      • Google, Facebook, Microsoft에 있으면 acceptance rate가 평균이상이었다.
      • "Theorem"이 들어있으면 score가 0.34정도 더 높다.
    • cf. [1] CMU/MIT/Cornell이면 deadline에서 1주이내에 arxiv에 올리면 score가 0.67 정도 더 높아진다... (a.k.a. arXiv effect)

[1] An Open Review of OpenReview: A Critical Analysis of the Machine Learning Conference Review Process - rejected from ICLR2021 yet got a positive review: "The paper should be brought to the attention of the various committees and made available somehow at the conference and acknowledged as a useful publication."

  • You Are the Best Reviewer of Your Own Papers: An Owner-Assisted Scoring Mechanism (NeurIPS 2021)

    • UPenn (단독저자...!)
    • Isotonic Mechanism을 제안
      • Input: reviewer들의 raw score, author들이 생각하였을때 자신이 제출한 paper들의 ranking
      • Output: adjusted scores (via solving a convex optim)
    • 완전히 theoretical paper!
    • 어떤 assumption하에, 저자들이 true ranking (i.e. 최대한 양심적으로)을 input으로 주어야 utility가 maximize된다는 것이 증명됨! (Theorem 1)
    • Future work:
      • paper의 개수도 variable
      • utility function의 heterogeneity (현재는 모든 paper가 똑같은 utility func을 쓰고 있지만, 어떤 paper는 "그냥" 내는 것일 수도 있고 어떤 paper는 "갈아서" 내는 것일수도 있다...!)
  • Parameter Prediction for Unseen Deep Architectures (NeurIPS 2021)

    • University of Guelph, Vector Institute, FMeta-AIR, McGill University
    • 현재 framework에선, 다른 network를 train하면서 얻은 정보를 전혀 사용하고 있지 않음!
    • ==> 큰 hypernetwork H를 다른 network에서 학습하여, 거기에 우리가 원하는 network를 넣어서 single pass!
      • input: network weights, architecture
      • output: optimized network weights
    • Summary
      • training (SGD, Adam..) 없이 model의 parameter를 예측!
        • ResNet-101, ResNet-152, Wide-ResNets, ViT...etc. & CIFAR10, CIFAR100, ImageNet,....etc.
        • CPU에서도 forward pass가 1초만에 가능함! (물론 train된 hypernetwork가 있다는 가정하에)
        • 실험들이 얼마나 좋은지는 제 수준에선 잘 모르겠네요 ㅠㅠㅠㅠ 도와주세요,,,
      • NAS에도 적용함!
      • 이걸 위해 새로운 dataset을 내놓음: DeepNets-1M
      • https://threadreaderapp.com/thread/1452813952719761416.html
    • Future work: 이렇게 나온 init을 theoretical하게 분석하는것도 상당히 흥미로울듯..? (저만 그런가요..)
      image

Extra

  • Deep Ensembles as Approximate Bayesian Inference
    • Bayesian DL 전문가이신 NYU의 Prof. Andrew Gordon Wilson의 blogpost
    • 결론: For the good of science, please stop calling deep ensembles a “non-Bayesian” competitor to standard approximate Bayesian inference procedures
    • 좀 더 정제된 결론:
      • There are no exact inference procedures in Bayesian deep learning.
      • Deep ensemble ~ numerical integration (feat. [1]) i.e. 이건 빼박 (approximate) Bayesian이다
      • Viewing deep ensembles as approximate Bayesian inference provides an actionable mechanism for improving the fidelity of approximate inference. For example, the MultiSWAG procedure [1], which performs Bayesian integration across and within basins of attraction, was directly inspired by this perspective and provides improved performance over deep ensembles.
    • 밑에 분노의 QnA를 보시면 재밌어요..! (특히 Bayesian에 관심 있으신분들 ㄱㄱ)
      image

[1] Bayesian Deep Learning and a Probabilistic Perspective of Generalization (NeurIPS 2020)

쓸데없는 논문

@nick-jhlee
Copy link

nick-jhlee commented Oct 31, 2021

STOC 2022 (ACM Symposium on Theory of Computing)의 author ordering에 대한 공고:
image

호오...
image

@nick-jhlee
Copy link

정부, 출입국 얼굴사진 1억7천만건 AI업체에 넘겼다

@nick-jhlee
Copy link

nick-jhlee commented Oct 31, 2021

상-당한 뒷북

The Nobel Prize in Physics 2021 was awarded "for groundbreaking contributions to our understanding of complex systems" with one half jointly to Syukuro Manabe and Klaus Hasselmann "for the physical modelling of Earth's climate, quantifying variability and reliably predicting global warming" and the other half to Giorgio Parisi "for the discovery of the interplay of disorder and fluctuations in physical systems from atomic to planetary scales."

  • 저 분 (Giorgio Parisi)의 spin glass theory 및 replica method (주변에 대학원 statistical physics 들은 분들께 여쭤보시면 뭔지 알려주실거에요 ㅎㅎ) 덕분에 70~90년대의 neural network theory에 상당한 영향을 끼침!
  • 다른 쓰이는 곳: combinatorial optimization, graphical model, complex network theory...etc
  • 원래 particle physics, quantum physics, astrophysics처럼 엄청 "순수"한 분야거나 semiconductors, microscope처럼 응용물리에서 항상 받아갔었는데, 이 statistical physics + network theory 분야에서 노벨상 나온건 처음이라 아주 역사적인 순간..

그냥 뭔가 이런 연구가 있었다는게 최신 deep learning과 large scale LM의 홍수에 잊혀지면 약간 슬플거 같애서 이렇게 소개해드려요,, (물론 저 홍수가 나쁘다는건 절대 아님다 ㅎㅎ)

@hollobit
Copy link

hollobit commented Oct 31, 2021

https://www.youtube.com/watch?v=4WQwW6FrDGc

ISO/IEC JTC 1/SC 42(AI), 10월 총회 개최 완료 (Virtual Moscow, 10/18 ~ 10/30)

https://www.iso.org/committee/6794475.html

  • 총 240여명의 참석자, 한국에서도 30명 넘게 참석
  • 1개 AG 신설 결정, 25건의 표준 개발 논의, 3건의 신규 표준 개발 결정 (data life cycle framework, transparency taxonomy, Treatment of unwanted bias in classification and regression machine learning tasks)

FDA's Good Machine Learning Practice for Medical Device Development: Guiding Principles

https://www.fda.gov/medical-devices/software-medical-device-samd/good-machine-learning-practice-medical-device-development-guiding-principles

  • 미국 식품의약국(FDA), 캐나다 보건부, 영국 MHRA(의약품 및 건강 관리 제품 규제 기관)는 GMLP(Good Machine Learning Practice) 개발에 도움이 되는 10가지 지침 원칙 공유
  • 향후 IMDRF(International Medical Device Regulators Forum), 국제 표준 기구 및 기타 협력 기관이 GMLP 드라이브 전망
  1. 전체 제품 수명 주기 동안 다분야 전문 지식 활용
  2. 우수한 소프트웨어 엔지니어링 및 보안 사례 구현
  3. 임상 연구 참가자 및 데이터 세트는 의도된 환자 모집단을 대표
  4. 훈련 데이터 세트는 테스트 세트 와 독립적
  5. 선택된 참조 데이터 세트는 최적의 사용 가능한 방법에 기반
  6. 모델 설계는 사용 가능한 데이터에 맞게 조정되고 장치의 의도된 사용을 반영해야
  7. 인간-AI ​​팀의 성과에 초점이 맞춰야
  8. 테스트 시 임상 관련 조건에서 장치 성능을 보여줘야
  9. 사용자에게 명확하고 필수적인 정보 제공
  10. 배포된 모델의 성능 모니터링 및 재학습 위험 관리

Kicking off NIST AI Risk Management Framework

https://www.nist.gov/news-events/events/2021/10/kicking-nist-ai-risk-management-framework

인공지능 전지역·전산업 확산 추진방향 수립 (10/29, 과기부, 4차산업혁명위원회 발표)

http://www.aitimes.com/news/articleView.html?idxno=141256
https://www.korea.kr/news/pressReleaseView.do?newsId=156477670

  • 과기정통부 17개 시·도 및 관계부처 합동 ‘인공지능 지역확산 추진방향’ 발표 -
  • 지역경제 재도약과 디지털 대전환 가속화를 위한 초광역 협업 사업(프로젝트) -
    · (혁신거점) 광주 집적단지를 국가 인공지능 혁신거점으로 고도화
    · (선도사업) 5대 권역별 강점을 반영한 대형 인공지능 사업(프로젝트) 기획
    · (특화융합) 지역별 주력산업 혁신을 위한 인공지능 융합 확대

스크린샷 2021-10-31 오후 5 04 00

The Coming Convergence of NFTs and Artificial Intelligence

https://www.coindesk.com/business/2021/10/25/the-coming-convergence-of-nfts-and-artificial-intelligence/

  • 3대 유형 : AI-generated NFTs, NFTs’ embedded-AI, AI-first NFT infrastructures

Expert warns that artificial intelligence could soon be able to 'hack' human beings

https://thehill.com/changing-america/respect/579196-expert-warns-that-artificial-intelligence-could-soon-be-able-to-hack
https://www.dailymail.co.uk/news/article-10147617/Israeli-historian-Yuval-Harari-warns-humans-cooperate-stop-human-hacking.html

  • Harari의 주요 주제 중 하나는 인공 지능 알고리즘이 인간의 삶에 대한 영향력을 높일 때 초래할 수 있는 위험한 변화에 대해 사람들에게 경고하는 것
  • 'Netflix는 우리에게 무엇을 봐야 하는지 알려주고 Amazon은 무엇을 사야 하는지 알려줍니다. 결국 10년, 20년 또는 30년 내에 그러한 알고리즘은 대학에서 무엇을 공부해야 하는지, 어디에서 일해야 하는지, 누구와 결혼하고 누구에게 투표해야 하는지 알려줄 수 있을 것입니다'라고 그는 말했습니다. 
  • 유일한 해결책은 데이터 수집을 규제하기 위해 국가들이 협력하는 것

Bruno Maisonnier, founder and CEO of AI firm AnotherBrain

  • We brought fire and people died from the fire, we brought nuclear and people died from that.
  • Each time we have the same reaction: First we fear and then we start to put the feedback and learn and put rules to get the positive out of this technology
  • The same goes with AI. The question is when do we have to set these rules?
  • Rules must be put but first, we must allow the evolution to happen.

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