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

[20210411] Weekly AI Arxiv 만담. #5

Closed
jungwoo-ha opened this issue Apr 9, 2021 · 9 comments
Closed

[20210411] Weekly AI Arxiv 만담. #5

jungwoo-ha opened this issue Apr 9, 2021 · 9 comments

Comments

@jungwoo-ha
Copy link
Owner

jungwoo-ha commented Apr 9, 2021

@tteon
Copy link

tteon commented Apr 9, 2021

  • Adaptive Filters and Aggregator Fusion for Efficient Graph Convolutions

  • adaptive filter 와 aggregator fusion 을 활용하여 parameter-efficient를 통해 좋은 성능ㅡmemory efficiency, lower latency , higher accuracy ㅡ을 보임.

  • 청자분들 대다수가 현업에 계시는 분들이라 생각되어 architecture improvement 관점과 다르게 파라미터 관점에서 접근한다는 점이 색달라 혹 도움이 되지 않을까 싶어 가져와보았습니다 :)

@sjYoondeltar
Copy link

sjYoondeltar commented Apr 11, 2021

AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control

  • 딥 미믹 저자 jason peng 이 공동 1저자로 들어간 논문

  • 시뮬레이션 내 캐릭터의 거동을 실제와 가깝게 재현하는 문제에 대해 data-driven method는 예전부터 제시가 많이 되어왔었습니다.

  • 그러나 모션 데이터로부터 클립 어노테이션을 해야한다던가 추가적인 모션 플래너를 달아야 하는 등의 번거로움이 있었습니다.

  • 논문에서는 캐릭터가 지정된 목표로 가는 태스크에 대한 reward와 함께, 모션 데이터셋 내의 움직임들의 스타일을 모방하기 위한 reward를 LSGAN을 사용하여 만들고, 해당 리워드 두개로 캐릭터를 학습시킵니다

  • 결과는 홈페이지, 유툽에서 나온대로 복잡한 태스크 수행과 스타일의 모방을 동시에 만족시키고 있습니다

  • https://xbpeng.github.io/projects/AMP/index.html

@nick-jhlee
Copy link

nick-jhlee commented Apr 11, 2021

Geometrically Enriched Latent Spaces <AISTATS 2021>
(이론하기 살짝 눈치...보여서 ㅎㅎ 뭔가 그래도 다들 재밌을 확률이 높은걸 가져와봤어용)

  • Bernhard Schölkopf 아저씨가 참여한 논문

  • generative model의 작동 방식이 latent space -> (nonlinear mapping) -> ambient space 이고, 보통 ambient space로는 Euclidean space(R^n)을 씁니다

  • 근데 저 nonlinearity땜에 생기는 distortion으로 인해 latent space가 ambient space를 정확히 나타내지는 못합니다

  • 이 논문에선 ambient space를 Riemannian manifold로 간주하고, 해당되는 Riemannian metric을 제시합니다.

  • 실험을 통해, interpretability가 더 높아졌고, data manifold 위에서 좀 더 "optimal"한 path가 선택됨을 볼 수 있습니다.

Capture

@qqueing
Copy link

qqueing commented Apr 11, 2021

@jshin49
Copy link

jshin49 commented Apr 11, 2021

Integrated Rolling Review system for NLP *ACL conferences
https://aclrollingreview.org/

@jshin49
Copy link

jshin49 commented Apr 11, 2021

위에는 링크가 없어서..
ViViT: A Video Vision Transformer

@jwlee-ml
Copy link

An Empirical Study of Training Self-Supervised Visual Transformers
Kaming He, 말 그대로 SSL 기반 transformer에 대한 다양한 실험이네요... 어느분이 읽어보시고 설명좀 부탁드립니다 ㅎㅎ

@rickiepark
Copy link

오늘 카패시가 트윗했던데 간단히 설명 부탁드려도 될까요.. :)
RepVGG: Making VGG-style ConvNets Great Again

@j-min
Copy link

j-min commented Apr 11, 2021

아까 Low-Complexity Probing via Finding Subnetworks 관련해서 언급했던 Supermasks in Superposition (NeurIPS 2020) 입니다.
https://mitchellnw.github.io/blog/2020/supsup/

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

9 participants