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

[20210314] Weekly Arxiv Casual Talk #2

Closed
jungwoo-ha opened this issue Mar 13, 2021 · 0 comments
Closed

[20210314] Weekly Arxiv Casual Talk #2

jungwoo-ha opened this issue Mar 13, 2021 · 0 comments

Comments

@jungwoo-ha
Copy link
Owner

jungwoo-ha commented Mar 13, 2021

image

  • Arxiv 논문 리스트
    • Multimodal Neurons in Artificial Neural Networks (https://openai.com/blog/multimodal-neurons/ )
      • CLIP 분석 통해 발견
      • We’ve discovered neurons in CLIP that respond to the same concept whether presented literally, symbolically, or conceptually.
      • 모달리티와 상관없이 동일 개념에 발화 --> 개념과의 연관성, Bias 이슈까지
      • 사람 인지에서의 개념과 추론등에서 일어나는 일과 연계가능할 것인가?
    • Neural 3D Video Synthesis ( https://neural-3d-video.github.io/ )
      • 비디오 멀티뷰 생성
    • SEER: https://arxiv.org/pdf/2103.01988.pdf
      • Facebook 에서 Instagram 1B로 학습한 self-supervised
      • random하고 curate 하지 않아도.. 심지어 데이터가 바뀌어도.. 성능에는 그다지..
      • 무한 데이터 공급원이 있고 모델이 충분히 크면 self-supervised로도 안정적 백본 학습 가능....
    • Rissanen Data Analysis: Examining Dataset Characteristics via Description Length (https://arxiv.org/abs/2103.03872 )
    • Im2Vec: Synthesizing Vector Graphics without Vector Supervision (https://arxiv.org/pdf/2102.02798.pdf )
    • Self-training Improves Pre-training for Natural Language Understanding (https://arxiv.org/abs/2010.02194 )
    • Fusion-in-Decoder code 공개 (https://github.com/facebookresearch/FiD )
    • Self-supervised Text-to-SQL Learning With Header Alignment Training (https://arxiv.org/pdf/2103.06402v1.pdf )
      • from 국민은행 + KAIST
      • NL-to-SQL을 위한 self-supervised / 테이블 header를 pseudo utterance로 하는 self-supervised NL-to-SQL 모델은 쉐어해서 Fine tune?
    • FAIRFIL: CONTRASTIVE NEURAL DEBIASING METHOD FOR PRETRAINED TEXT ENCODERS (https://arxiv.org/pdf/2103.06413v1.pdf )
      • from Duke U (ICLR 2021)
      • Pretrained LM 기반 sentence-level debiasing 연구 --> 편향성 제어에 도움?
    • Metapaths-guided Neighbors-aggregated Network for Heterogeneous Graph Reasoning (https://arxiv.org/pdf/2103.06474v1.pdf )
      • from Alibaba
      • user, item들이 마구 섞여 있는 hetero graph를 metapath를 통해학습 (node embedding활용) 하고 node, edge 분류하는..
      • 실제 online A/B도 해봤다고 (역시 알리바바)
    • Does the Magic of BERT Apply to Medical Code Assignment? A Quantitative Study (https://arxiv.org/pdf/2103.06511v1.pdf )
      • from Aalto Univ.
      • MIMIC-III 류의 의료기록데이터에서 BERT류의 LM 가능성 연구.
    • WenLan: Bridging Vision and Language by Large-Scale Multi-Modal Pre-Trainin (https://arxiv.org/abs/2103.06561v1 )
      • 대륙의 CLIP, 회사는 아니고 학교에서 나옴. 저자 엄청 많음
      • CLIP과 마찬가지로 백본 학습 후 I2T retrieval 같은 곳에 적용
    • M6: A Chinese Multimodal Pretrainer (https://arxiv.org/pdf/2103.00823.pdf )
      • 대륙의 DALL-E
      • Alibaba와 칭화대의 합작품. 알리바바 소유의 초대량 데이터를 활용해 만든 생성모델
      • DALL-E가 대략 12B 인데 비해 M6는 100B까지
    • Privacy-preserving Object Detection (https://arxiv.org/abs/2103.06587v1 )
      • from U. of Oxford
      • Object detection을 위한 훈련데이터에서 사람 개인정보에 해당하는 부분을 날리면 성능감소가 생기나?
      • StarGAN활용해서 원본사람 얼굴 바꿔넣고 (대충), Blur 도 해보고 다양한시도
      • 결론적으로 detection 성능이 크게 영향을 받는 것은 아닌 것 처럼 보임 (적어도 COCO에 대해서는)
    • MagFace: A Universal Representation for Face Recognition and Quality Assessment (https://arxiv.org/pdf/2103.06627v1.pdf )
      • CVPR 2021 Oral
      • 새로운 loss 도입으로 퀄이 좋아 쉬운 샘플은 embedding space class center에 가깝게하고 퀄이 나쁘고 noisy한 이미지는 저멀리. Quality를 임베딩에서 동시에 평가 가능한 신박한 방법
      • https://github.com/IrvingMeng/MagFace
    • ENTRUST: Argument Reframing with Language Models and Entailment (https://arxiv.org/abs/2103.06758v1 )
      • Columbia U. & Google
      • 논의에서 긍/부정과 관련된 프레이밍 연구. Framing과 관련된 dataset. 악성댓글의 순화 같은 데다 적용가능?
      • Code와 데이터는 coming soon
    • Automatic Social Distance Estimation From Images: Performance Evaluation, Test Benchmark, and Algorithm (https://arxiv.org/pdf/2103.06759v1.pdf)
      • 핀란드, 덴마크 연구
      • 제목 그대로 사회적 거리두기 체크 AI를 위한 데이터, 평가 프로토콜을 제안함
      • 빠른 도입이 시급해 보임
    • Sampling methods for efficient training of graph convolutional networks: A survey (https://arxiv.org/pdf/2103.05872v1.pdf )
      • 최근 추천 시스템의 대세 GCN
      • 문제는 그래프가 엄청 크면 (보통 인더스트리에서..) node, edge가 너무 많아 학습이 어려움
      • 원래 그래프 성질과 유사하도록 sampling 을 잘하는게 중요한데 이와 관련된 sampling method의 survey 연구
    • VideoMoCo: Contrastive Video Representation Learning with Temporally Adversarial Examples (http://arxiv.org/abs/2103.05905v1 )
      • from 텐센트
      • MoCo의 video 버전, self-supervised temporal feature representation.
      • 보통 video data aug에서도 (VideoMix) temporal 정보 중요도가 떨어지는데 (spatial만 봐도 되는 게 구조때문인지 데이터 때문인지 몰겠지만) 여기선 temporal을 강제로 보게하는 여러 트릭이 들어가는 듯
      • UCF101, HMDB51 에 대해서 해봤다고...
    • Pretrained Transformers As Universal Computation Engines( https://arxiv.org/abs/2103.05247v1 )
      • 자연어데이터로 pretraining 후에 다른(예: image) data로 finetuning
    • Fast and Accurate Model Scaling (https://arxiv.org/abs/2103.06877v1 )
      • by R. Girshick, FAIR (CVPR 2021)
      • CNN 모델 구조 설계를 위한, depth, width, resolution등 다양한 스케일링 관련 실험
      • CNN architecture 관심 많으신분들께는 매우 유용할 각..
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

1 participant