Learning to summarize from human feedbackharmonylab
公開URL:https://arxiv.org/abs/2009.01325
出典:Nisan Stiennon, Long Ouyang, Jeff Wu, Daniel M. Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, Paul Christiano : Learning to summarize from human feedback, arXiv:2009.01325 (2020)
概要:言語モデルが強力になるにつれて、モデルの学習と評価は特定のタスクで使用されるデータとメトリクスによってボトルネックになることが多い。要約モデルでは人間が作成した参照要約を予測するように学習され、ROUGEによって評価されることが多い。しかし、これらのメトリクスと人間が本当に気にしている要約の品質との間にはズレが存在する。本研究では、大規模で高品質な人間のフィードバックデータセットを収集し、人間が好む要約を予測するモデルを学習する。そのモデルを報酬関数として使用して要約ポリシーをfine-tuneする。TL;DRデータセットにおいて本手法を適用したところ、人間の評価において参照要約よりも上回ることがわかった。
SIGGRAPH 2017 Theater: SIGGRAPH in Japanese and Japan CG Showcase by Yoichi O...Yoichi Ochiai
SIGGRAPH 2017 Theater: SIGGRAPH in Japanese and Japan CG Showcase by Yoichi Ochiai
Digital Nature Group presented works in 2 sessions on 10 publications at SIGGRAPH 2017. See you in Los Angeles.
http://digitalnature.slis.tsukuba.ac.jp/2017/07/dng-siggraph-2017/
https://youtu.be/xc-ChP_Tfk0?t=51m35s
Learning to summarize from human feedbackharmonylab
公開URL:https://arxiv.org/abs/2009.01325
出典:Nisan Stiennon, Long Ouyang, Jeff Wu, Daniel M. Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, Paul Christiano : Learning to summarize from human feedback, arXiv:2009.01325 (2020)
概要:言語モデルが強力になるにつれて、モデルの学習と評価は特定のタスクで使用されるデータとメトリクスによってボトルネックになることが多い。要約モデルでは人間が作成した参照要約を予測するように学習され、ROUGEによって評価されることが多い。しかし、これらのメトリクスと人間が本当に気にしている要約の品質との間にはズレが存在する。本研究では、大規模で高品質な人間のフィードバックデータセットを収集し、人間が好む要約を予測するモデルを学習する。そのモデルを報酬関数として使用して要約ポリシーをfine-tuneする。TL;DRデータセットにおいて本手法を適用したところ、人間の評価において参照要約よりも上回ることがわかった。
SIGGRAPH 2017 Theater: SIGGRAPH in Japanese and Japan CG Showcase by Yoichi O...Yoichi Ochiai
SIGGRAPH 2017 Theater: SIGGRAPH in Japanese and Japan CG Showcase by Yoichi Ochiai
Digital Nature Group presented works in 2 sessions on 10 publications at SIGGRAPH 2017. See you in Los Angeles.
http://digitalnature.slis.tsukuba.ac.jp/2017/07/dng-siggraph-2017/
https://youtu.be/xc-ChP_Tfk0?t=51m35s
Exponential Organizations - Why new organizations are 10x better, faster and ...Yuri van Geest
Exponential Organizations (ExOs, #ExponentialOrgs) - authored by Yuri van Geest, Salim Ismail, Peter Diamandis and Mike Malone and published by Singularity University Press - how to build exponential organizations with exponential technologies and new organizational techniques for an exponential era.
This is first book integrating all key organizational and technology trends into a new and holistic 11 attribute framework applicable for startups, mid markets and corporates. To create exponential organizations instead of classic, linear ones which were developed more than 100 years ago.
We already received the Best Business Book of the Year 2014 Award by Frost & Sullivan and are accepted in the prestigeous C-Suite Book Club.
The book has been thoroughly researched in the last 30 months and we looked for patterns in the most important exponentials companies in the world in the last 6 years like Waze, Tesla, Airbnb, Uber, Xiaomi, Netflix, Valve, Google (Ventures), GitHub, Quirky and 60 other companies including successful corporates like GE, Haier, Coca Cola, Amazon, Citibank and ING Bank. We interviewed 70 global leaders and thinkers like Marc Andreessen, Arianna Huffington, Steve Forbes, Philip Rosedale, Tim O'Reilly, Chris Anderson and many others.
The book is already an Amazon bestseller in the pre-order phase since June, 2014 in the categories Startups, Business Management and Innovation.
This slide shows (1) AI and Accountability , (2) AI Ethics, (2) Privacy Protection. Several AI ethics documents such as IEEE EAD, EC-HELG Ethics Guideline for Trustworthy AI, Social Principles of Human-Centric AI(Japan), focus on AI's transparency, accountability and trust. We follow the discussions of these documents around the above (1),(2) and (3) topics.
What is Accountability of AI? We answer to this question by clarifying responsibility, explainability and liability of limited autonomous AI with several bright and dark real examples.
Then we move to the concept of "Trust " which is of not limited to single AI system but group AI ‘s behavior.
K-anonymization has been regarded as a great method to make a bad person indistinguishable among k people whose quasi identifiers are same.
It, unfortunately, has a problematic side effect of defamation. In this case, defamation means the case where other good k-1 people are suspected as a bad person because both of a bad person and good people have the same quasi identifiers because of k-anonymization. This slide shows a mathematical model of defamation and proposes an algorithm which minimizes the probability of defamation.
Social Effects by the Singularity -Pre-Singularity Era-Hiroshi Nakagawa
Contents:
Stance of scientists community against Pre-Singularity problems
Amplification vs. Replacement
AI takes over jobs
Boarder line between amplification and replacement
Autonomous driver: trolley problem
The right to be forgotten
Towards black box
Responsibility
Vulnerability of financial dealing system made of many AI agent traders connected via internet
AI and weapon
Filter bubble phenomena
Analogy: Selfish gene
AI and privacy
The right to be forgotten, Profiling and Don’t Track
Feeling of friendliness to android
Again self conscious and identity