John Schulman - Reinforcement Learning from Human Feedback: Progress and Challenges

推荐理由

There's a very good talk by John Schulman, chief architect of ChatGPT

推荐KKKKkkkk1 · 原文2023年5月

说明

这是2023年由UC Berkeley EECS发布的一部演讲，主题归在Machine learning & deep learning。它讲的是rlhf for language models这一支里的具体问题，而不是把整个领域扫一遍。适合已经能看懂基本术语、想继续往下走的人。内容上可以概括成：UC Berkeley EECS colloquium by OpenAI's John Schulman on reinforcement learning from human feedback. He covers why RL helps with truthfulness, hallucination, and the limits of behavior cloning.这些读者给出的理由，核心是：Recommended in LLM threads as an authoritative explanation from a ChatGPT architect of why models hallucinate and why imitation is not enough.它贴着当时的技术，看的时候要记得那是那个年份的判断。从2023年到2023年，有2位不同的读者在2个不同月份把它推荐出来，其中明确叫别人去看的有0次。适合一次性看完。讲者把一个判断讲透，演示是在给这个判断配例子。如果已经会了基本定义，就把注意力放在它怎么把rlhf for language models变成能跟着复述的例子。

视频信息

频道

UC Berkeley EECS

时长

1 小时 3 分

发布年份

2023

同主题更多

• The spelled-out intro to neural networks and backpropagation: building micrograd
• But what is a neural network? | Deep learning chapter 1
• AlphaGo - The Movie | Full award-winning documentary
• Transformers, the tech behind LLMs | Deep Learning Chapter 5
• USENIX Security '18-Q: Why Do Keynote Speakers Keep Suggesting That Improving Security Is Possible?
• Lecture 1 | Machine Learning (Stanford)