Skip to content

[20230702] Weekly AI ArXiv 만담 시즌2 - 22회차 #88

@jungwoo-ha

Description

@jungwoo-ha

상반기 회고 1마디씩

News

  • Conferecnes

    • ACL 2023 (7.9 - 14, Toronto) - 네이버 논문 10개 발표 많이 들러 주세요~
  • "사용료 내느니 뉴스 뺄래"... 구글·메타, 캐나다서 뉴스 제공 중단 선언

  • "초거대AI 보유는 핵보유와 같은 경쟁력"...네이버·LG 등 초거대AI 협의회 출범(종합)

  • ChatGPT vs. Bard의 요즘 근황
    image

  • Tortoise Global AI Index 2023

    • Stanford AI Indext Report 와 더불어 가장 많이 활용되는 Tortoise media의 AI 인덱스 보고서 2023년 버전 업데이트
    • 항목별 설명
      • Talent focuses on the availability of skilled practitioners in artificial intelligence solutions.
      • Infrastructure assesses the reliability and scale of access infrastructure, from electricity and internet to supercomputing capabilities.
      • Operating Environment focuses on the regulatory context and public opinion on artificial intelligence.
      • Research looks at the extent of specialist research and researchers, including numbers of publications and citations in credible academic journals.
      • Development focuses on the development of fundamental platforms and algorithms upon which innovative artificial intelligence projects rely.
      • Government Strategy gauges the depth of commitment from national governments to artificial intelligence; investigating spending commitments and national strategies.
      • Commercial focuses on the level of startup activity, investment and business initiatives based on artificial intelligence.
    • 한국은 작년 7위에서 6위로 한계단 상승
    • 이번에는 랭킹외에 상대평가 점수도 함께 제공: 미국과 중국의 차이가 엄청나게 큼 (100점 vs. 61.5점)
    • 싱가폴이 3년 연속 약진해서 3위까지 오름. 특히 규모감은 약해도 intensity에서 1위를 차지. 특히 Commercial이 눈에 띔
    • 우리나라는 반대로 인재에서는 많이 올라왔지만 Commercial이 상대 점수로도 너무 약함. 스타트업과 투자 생태계 재점검이 필요
    • 아래쪽에 항목별 x-y, 색 설정을 통한 시각화 툴로 재미있는 분석 가능
      image
      image
      image
      image

ArXiv

  • Generate Anything Anywhere in Any Scene

    • Text-to-image 에서 개인 물체, 크기, 위치 모두 잘 컨트롤되도록 만든 연구 (Dreambooth 와 GLIGEN 단점 커버): U of Wisconsin-Madison 이용재 교수님 연구실
    • Training time에 크기 위치 data augmentation 기법을 활용해서 object identity 포함 disentanglement 문제해결
    • GLIGEN처럼 inference에서 위치, 크기 컨트롤 위한 GLIGEN adapter layer를 넣음
    • 프로젝트 페이지: https://yuheng-li.github.io/PACGen
      image
      image
      image
  • How Long Can Open-Source LLMs Truly Promise on Context Length?

    • LongChat-13B & 7B-16k 성능에 대한 기술 블로그
    • Rotary embedding 에 대화데이터 FT 스타일
    • 실제로 Open-source LLM들이 long context 의 retrieval task에서 생각보다 강력하지 못함 (MPT 시리즈 특히)
      • Coarse-grained topic retrieval, Fine-grained line retrieval 모두에서 어려움
      • 그래도 Claude나 GPT-3.5 Turbo는 상당히 능력을 보여줌
    • Positional interpolation 은 유심히 뜯어볼만 해보임
      image
      image
      image

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions