Skip to content
New issue

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

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

Already on GitHub? # to your account

[20220710] Weekly AI ArXiv 만담 - 58회차 #58

Closed
jungwoo-ha opened this issue Jul 9, 2022 · 6 comments
Closed

[20220710] Weekly AI ArXiv 만담 - 58회차 #58

jungwoo-ha opened this issue Jul 9, 2022 · 6 comments

Comments

@jungwoo-ha
Copy link
Owner

jungwoo-ha commented Jul 9, 2022

News

ArXiv

@veritas9872
Copy link

veritas9872 commented Jul 10, 2022

Mitigating Neural Network Overconfidence with Logit Normalization:
Arxiv: https://arxiv.org/abs/2205.09310
GitHub: https://github.com/hongxin001/logitnorm_ood

ICML 2022에 Accept된 논문입니다. 기존 Neural network에서 over-confident prediction의 원인에 대한 많은 분석이 있었는데 본 논문에서는 softmax를 통과하기 전의 logit 값의 norm을 constant하게 fix함으로써 overfitting을 방지할 수 있다고 합니다.
ResNet 및 CNN 구조와 CIFAR100 등 데이터셋에 실험을 진행해 아직 large-scale confirmation이 진행되지는 않았지만 구현 방법이 매우 단순하며 기존 보았던 last layer gradient/norm explosion 이슈와 관련이 있을 것으로 예상됩니다.

image

YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors:
Arxiv: https://arxiv.org/abs/2207.02696
GitHub: https://github.com/jinfagang/yolov7
image

YOLOv7이 공개되었습니다. 기존의 방법들의 여러 테크닉을 조합하고 새로운 repository에서 detection 뿐만 아니라 instance segmentation 및 transformer backbone 등 방법론적, 기술적 novelty가 추가되었는데 real-time detection 및 segmentation을 하시는 분들께 도움이 될 것으로 생각됩니다.

Perspective (In)consistency of Paint by Text:
Arxiv: https://arxiv.org/abs/2206.14617

최근 이목을 끈 Parti, Imagen 등 모델에서 원근법의 불일치를 분석한 논문입니다. 인간의 시각적 감각은 원근법이 불일치하더라도 인지하는데 어렵기 때문에 생성모델에서도 그런 특성이 반영된 것으로 보입니다. 평행선의 소실점 불일치 현상을 관찰하면 진짜 영상인지 생성된 영상인지 파악하는데 사용할 수 있음을 보여줍니다.

image

흥미로운 연구:

PrefixRL: Optimization of Parallel Prefix Circuits using Deep Reinforcement Learning: https://arxiv.org/abs/2205.07000

EfficientFormer: Vision Transformers at MobileNet Speed:
Arxiv: https://arxiv.org/abs/2206.01191
GitHub: https://github.com/snap-research/EfficientFormer

@hollobit
Copy link

hollobit commented Jul 10, 2022

Meta가 자사 번역 AI 모델 'NLLB-200'와 'FLORES-200'을 공개

https://about.fb.com/news/2022/07/new-meta-ai-model-translates-200-languages-making-technology-more-accessible/
https://ai.facebook.com/blog/teaching-ai-to-translate-100s-of-spoken-and-written-languages-in-real-time/
https://github.com/facebookresearch/flores

  • 'No Language Left Behind(NLLB)'
  • 200가지 언어를 기존 AI 연구 대비 평균 44% 향상된 품질로 번역할 수 있다고
  • NLLB-200 모델과 FLORES-200 데이터 세트, 모델 교육 코드와 교육 데이터 세트를 다시 만들기 위한 코드도 공개
  • 국제협력 연구들을 위해 200,000달러 지원금 지원 및 아프리카 언어 번역을 위한 연구원 모집

‘해킹도 ML로 진화 중’ 머신러닝 해킹 수법 9가지

https://www.ciokorea.com/news/240019#csidx637920ef9641af3a010003ae99b44ee

  1. 스팸 필터를 속이는 스팸 메일
  2. 더 정교해진 피싱 메일
  3. 향상된 패스워드 해킹
  4. 딥 페이크 사기
  5. 기성 보안 도구 우회
  6. 공격 상대 정찰
  7. 자동화 에이전트
  8. AI 오염 공격
  9. AI 퍼징(AI fuzzing)

Practical Attacks on Machine Learning Systems

https://research.nccgroup.com/2022/07/06/whitepaper-practical-attacks-on-machine-learning-systems/

  • NCC Group 수석 과학자 Chris Anley 작성
  • 민감한 데이터를 사용하여 ML 시스템을 교육하는 것은 근본적으로 안전하지 않은 것으로 보임
  • 신경망 분류기는 적대적 교란 공격에 의해 쉽게 오분류될 수 있다는 점에서 "취약한" 것으로 보임
  • 원격 공격자가 모델을 실행하고 결과를 관찰하여 훈련된 모델의 사본도 추출할 수 있음

기계 학습 및 자연어 처리 기술를 이용하여 전쟁으로 피해를 입은 인프라를 식별

https://www.undp.org/blog/ukraine-machine-learning-algorithms-and-big-data-scans-used-identify-war-damaged-infrastructure
https://tinyurl.com/semdam

  • ACLED 데이터를 활용하여 전쟁 정보 수집하고, 다양한 정보들을 분류, 통합 분석

[김 소장의 ‘핏(FIT)’] 고인의 모습과 고인의 목소리, 당신의 추억 방법은?

https://it.donga.com/102456/

AI 기술을 활용한 ‘디지털 불멸’ 산업의 성장... 여러분은 어떻게 생각하시나요 ?

Temporal quality degradation in AI models

https://www.nature.com/articles/s41598-022-15245-z

  • AI "노화"에 대한 첫번째 분석 시도
  • AI 노화란 마지막 모델 훈련 이후 시간이 지날수록 AI 모델 품질이 저하되는 복잡하고 다면적인 현상
  • 4가지 다른 산업(의료 운영, 운송, 금융 및 날씨)의 데이터 세트와 4가지 표준 기계 학습 모델을 사용하여 주요 시간 저하 패턴을 식별
  • 일시적인 품질 저하의 잠재적 원인을 나타내고 노화를 감지하고 그 영향을 줄이는 접근 방식을 제안

@nick-jhlee
Copy link

https://twitter.com/leonderczynski/status/1545133479385505792?s=21&t=BbD99qOp96vvkA2w5-h1zg

True story?

  1. best paper 달라고 말한게 아니라 best paper 후보 선정되니까 내이름 껴달라고함
  2. NeurIPS 2021때 AC들끼리 채팅방이 있었는데, 덕분에(?) 논문 12편됨
    a. 8 people이 있는 wechat에서 각자의 paper를 bidding하겠다고 함
    b. 실제로 그렇게 함
  3. 흔하답니다: collusion ring

@nick-jhlee
Copy link

nick-jhlee commented Jul 10, 2022

필즈상 2022

  • June Huh! (combinatorics + algebraic geometry)
  • James Maynard (analytic number theory)
  • Maryna Viazovska (number theory, sphere packing)
  • Hugo Duminil-Copin (probability theory, percolation theory)

@nick-jhlee
Copy link

nick-jhlee commented Jul 10, 2022

급홍보 (for theory-oriented people)

  • https://mltheory-korea.github.io
  • 현재 청강 가능한 statistical physics + deep learning theory 세미나를 제가 organize하고 있습니당 (관심 있으시면 연락주세요 ㅎㅎ)

@kimyoungdo0122
Copy link

kimyoungdo0122 commented Jul 17, 2022

News

  • 진원님 모두팝 행사!!

    • 다음주 화요일(19일) 모두의연구소 강남캠퍼스에서 오프라인으로 진행!
    • 주제는 '시간 순으로 ImageNet SOTA 모델 살펴보기'입니다
      image
    • 신청 링크
  • Microsoft Responsible AI Standard v2

    • 19년에 발표한 Responsible AI standard보다 더 실제적으로, 실무적으로 따를 수 있는 지침 대폭 추가
    • Accountability(책임성), Transparency(투명성), Fairness(공정성), Reliability and Safety(신뢰성과 안전성), Privacy and Security(개인정보 보호와 보안), Inclusiveness(포용성)
    • 각 항목들마다 문서화, 원칙, 평가 등 구체적으로 실행 가능한 가이드가 많이 추가되었습니다
    • 실제로 AI로 서비스를 개발 및 배포하는 조직에서 몇몇 항목들 참고하시면 좋을 듯합니다(기획자나 매니저가)
  • Greater creative control for AI image generation

    • 논문링크
    • Text2image 에서 기존의 텍스트 prompt에서 스케치와 같은 모델 결과물 제어에 필요한 constraint를 추가하는 것에 대한 연구
    • 올해 ECCV에서 oral 프레젠테이션 한다고하네요!
      image
      image

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

No branches or pull requests

5 participants