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

[20210620] Weekly AI ArXiv 만담 #14

Closed
jungwoo-ha opened this issue Jun 19, 2021 · 4 comments
Closed

[20210620] Weekly AI ArXiv 만담 #14

jungwoo-ha opened this issue Jun 19, 2021 · 4 comments

Comments

@jungwoo-ha
Copy link
Owner

jungwoo-ha commented Jun 19, 2021

@nick-jhlee
Copy link

nick-jhlee commented Jun 20, 2021

  • Very Deep Graph Neural Networks Via Noise Regularisation
    • DeepMind (Graph Representation의 rising 스타, Petar Velickovic 참여...!)
    • GNN을 deep하게 마구마구 쌓으면 여러 문제가 생김! -> oversmoothing (Chen et al., AAAI'20), bottleneck from over-squashing (Alon & Yahav, ICLR'21)
      • ANN에서도 너무 deep하게 쌓으면 문제가 많음 -> gradient vanishing/explosion, slow training...etc.
    • Noisy node regularisation을 제안함:
      • input graph의 (node, edge, graph attribute)를 perturb함
      • graph property prediction이면 autoencoding loss 추가 (~ denoising autoencoder, DAE)
      • ANN에서 Gaussian noise injection과 상당히 유사함
    • " Our results show this regularisation method allows the model to monotonically improve in performance with increased message-passing steps."
    • "We demonstrate that graph networks of large depth outperform shallower models even on small graphs, such as QM9, and can lead to substantial gains on very difficult datasets such as OC20. "
    • 여기서도 정확히 어떤 regularisation 효과가 있는지 수학적으로 밝히는게 하나의 future work...?
      • 실제로 ANN에서 Gaussian noise injection의 경우에는, Fourier domain에서 high frequency를 penalize하는 효과가 있다는 것이 밝혀짐 (Camuto et al., NeurIPS'20)

  • Do Transformers Really Perform Bad for Graph Representation?
    • Dalian University of Technology, Princeton, Peking, Microsoft Research Asia
    • GNN에서 transformer가 생각보다 SOTA를 잘 못 achieve 함... (다 잘하면 사기캐긴하죠)
    • 그래서. Graphormer 등-장
    • "Centrality Encoding", "Spatial Encoding"

  • Does Knowledge Distillation Really Work?
    • NYU (Bayesian Deep Learning의 rising 스타, AGW 교수님 참여...!), Google Research
    • KD 잘 되는건 모두가 아는 사실. BUT, 이게 진짜 우리가 생각한대로 작동하는걸까??
    • student-teacher agreementgood generalization of student를 분리함!
    • "student가 teacher를 배운다"가 맞는건가...?
      • NOPE! 실제로 independently trained model과 다를바가 없을 정도로 student와 teacher 사이의 agreement가 매우 낮다...!
    • student가 teacher를 배울 만큼 capacity가 없는건 아닌가?
      • NOPE! 아무런 상관이 없고, 실제로 self-distillation의 success가 있으려면 student가 반항아가 되어야함
    • (기타 등등)
    • 결론:

  • FedBABU: Towards Enhanced Representation for Federated Image Classification
    • KAIST (저를 뽑아주신 윤세영 교수님 참여...!)
    • FL에는 두 방향이 있음: a model that works for everyone decently or models that are personalized for each individual
    • 하지만 FedAvg에서 보니, 서로 충돌함...! (better global => no improvement in personalization)
    • Why?? head (classifier)를 각자 다르게 training을 해서 문제가 생김!!
    • Solution: head를 절대 training 시키지 말고 random하게 놔둔다. 즉 body만 update!
      • 각 client마다, classifier 전까지만 training을 시켜서 실제적으로 representation만 바꾸는 것!
      • Representation learning에서 비슷한 방향성을 가지고 ICLR'21에서 BOIL을 제안했었음...! (물론, BOIL에선 outer loop에서 classifier update를 하지만, FedBABU에선 거의 고정)
    • "이거의 장점은, client들이 서로 다른 data구성을 가지는 상황에서 representation을 학습하는 것이 제각각이 될 위험이 있는데 이건 고정된 기준점을 가지고 학습을 하기에 좀 더 효과적으로 배울 수 있게 된다."
    • " Extensive experiments show consistent performance improvements and an efficient personalization of FedBABU."

@veritas9872
Copy link

Knowledge distillation: A good teacher is patient and consistent
https://arxiv.org/abs/2106.05237
Google Research에서 Knowledge Distillation을 사용할 때 어떤 요소가 중요한지 분석해준 실험 논문입니다.
Knowledge Distillation은 많이 사용되는 만큼 어떻게 할때 잘 되는지 알면 도움이 많이 될 것 같습니다.

Efficient Deep Learning: A Survey on Making Deep Learning Models Smaller, Faster, and Better
https://arxiv.org/abs/2106.08962
제목 그대로 딥러닝을 더 효율적으로 학습하고 inference하는 방법들을 한군데에 정리한 Google Research의 리뷰 페이퍼입니다.
대부분 한번쯤 들어보셨을 내용이지만 한군데에 정리되어 있다는게 도움이 되어 공유합니다.

@hollobit
Copy link

4차 산업혁명 시대, 인공지능(AI)과 빅데이터 분야 표준화 시급 - 국가기술표준원

보건의료 데이터‧인공지능 혁신전략(’21~’25년) 수립 - 보건복지부 (6/3)

G7 보건의료 관계장관회의 성명 요약 (6/4)

AI can now convincingly mimic cybersecurity experts and medical researchers

인공지능 육성 및 신뢰 기반 조성 등에 관한 법률안 입법공청회 (6/18)

@hollobit
Copy link

Data-Centric AI Competition
https://https-deeplearning-ai.github.io/data-centric-comp/

# 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

4 participants