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

[20210905] Weekly AI ArXiv 만담 #23

Closed
jungwoo-ha opened this issue Sep 4, 2021 · 3 comments
Closed

[20210905] Weekly AI ArXiv 만담 #23

jungwoo-ha opened this issue Sep 4, 2021 · 3 comments

Comments

@jungwoo-ha
Copy link
Owner

jungwoo-ha commented Sep 4, 2021

@hollobit
Copy link

hollobit commented Sep 5, 2021

“AI옵스(AIOps)가 대세”…IDC, 2023년 기업 75%, AIOps 채택 전망

Facebook Apologizes After A.I. Puts ‘Primates’ Label on Video of Black Men

Only Humans, Not AI Machines, Get a U.S. Patent, Judge Says

The term AI overpromises. Let's make machine learning work better for humans instead

How open-source software shapes AI policy

  1. OSS SPEEDS AI ADOPTION
  2. OSS HELPS REDUCE AI BIAS
  3. OSS AI TOOLS ADVANCE SCIENCE
  4. OSS AI HELPS AND HINDERS TECHNOLOGY SECTOR COMPETITION
  5. OSS CREATES DEFAULT AI STANDARDS

AI Weekly: An outline for government regulation of AI

Sanas aims to convert one accent to another in real time for smoother customer service calls

Understanding, explaining, and utilizing medical artificial intelligence (Nature Human Behaviour)

AIMe – A standard for artificial intelligence in biomedicine

Maastricht University (UM)를 비롯한 여러 대학의 국제 연구에서 생물 의학 분야의 인공 지능(AI) 작업에 대한 표준화된 레지스트리를 제안

@ghlee3401
Copy link

ghlee3401 commented Sep 5, 2021

Paper

  • Neural HMMs are all you need (for high-quality attention-free TTS)

    • Sample page : https://shivammehta007.github.io/Neural-HMM/
    • Task : Text to Speech (TTS)
    • Motivation
      • neural seq2seq TTS가 HMM을 이용한 parametric 방법보다 좋다고 알려져 있지만, 이론적으로 설명가능한 확률 모델도 아니고, attention을 위해서 학습 시간이 길어지는 등의 문제가 있음
      • HMM과 neural TTS의 장점을 결합하여 두 개의 장점을 모두 취하는 방법을 제안
      • Tacotron 2의 attention을 neural network를 이용한 hidden-Markov model로 대체
    • Contribution
      • 적은 파라미터를 가지면서도 성능이 많이 떨어지지 않음
      • attention이 필요 없고 align을 빠르게 배움
      • HMM 기반이 neural TTS 수준의 음질을 보여준 것이 처음

image

  • You Only Hear Once: A YOLO-like Algorithm for Audio Segmentation and Sound Event Detection

    • Task : Audio segmentation and sound event detection
    • Contribution
      • 기존 SOTA인 CRNN 보다 RNN을 사용하지 않아 빠르면서도 좋은 성능을 보임
      • segmentation이나 event detection 문제를 해결하기 위하여 기존 방법들이 각 프레임에 대하여 클래스를 예측하는 문제였지만, 시간 경계를 직접 regression으로 예측하기 때문에 후처리가 필요하지 않음
    • Method
      • MobileNet architecture를 수정하였고 CNN으로 이루어짐
      • input으로는 mel-spectrogram 사용
      • output은 각 6 x #frames 음성 혹은 뮤직이 있는 구간인지 예측 (classification) 하고 시작시간과 끝 시간을 예측 (regression)
      • 샘플을 8s로 나누고, 다시 26개로 나누어 0.307로 나누어 label을 만듦
      • 예를 들어 music의 경우 시작 : 0.65 * 0.307 = 0.2 (s), 끝 : 130.307 + 0.9750.307 = 4.29(s)

image

@Kyung-Min
Copy link

Kyung-Min commented Sep 5, 2021

  • Paper
    • Graph Attention Multi-Layer Perceptron

      • Scalable GNN을 만들기 위한 시도
      • 모델 크기를 키울 때 노드 피처를 계산하는 시간이 오래 걸림. 매 hop 마다 샘플링하고 receptive field size 커지고.. 그래서 노드 피처를 미리 다 뽑아놓고, mlp를 통해 feature transformation을 시킴
      • 기본적인 구조는 SimpleGCN과 비슷하지만, 여러개의 hop에 걸쳐 얻은 피처들을 전부 활용하는 것에 차이가 있음
      • Adjacent matrix의 K 거듭제곱을 통해 hop 별로 노드의 feature를 미리 다 뽑아 놓고, attention을 통해 feature 통합
      • 그리고 MLP를 통해 transformation
        image
      • ogbn-products과 ogbn-papers100M에 테스트
    • AI 가 발전하는 방향 from Foundation 모델

      • image
      • emergence와 homogenization를 증가시키는 방향으로 발전
      • 전문가시스템 (룰베이스) -> 기계학습 -> 딥러닝 -> 파운데이션 모델
      • 전문가시스템은 emergence가 없음. 프로그래머가 작성한 룰 그대로 시스템 동작
      • 기계학습을 사용하면 how가 생김. 머신러닝을 통해 입출력 사이 관계가 학습되더라 (예> f=ma)
      • 딥러닝을 사용하면 features가 생김. CNN을 ImageNet에 학습시켰더니 이미지 feature를 뽑을 수 있게되더라
      • 파운데이션 모델을 사용하면 functionalities가 생김. GPT-3를 학습하니 시도 쓰고, 코딩도 하고, 마케팅 문구도 쓰더라

# 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