#14 Aligning AI models for healthcare | Understanding Reinforcement Learning from Human Feedback (RLHF)

How do we align AI models for healthcare? 👨‍⚕️ And importantly, the moral codes and ethics that we practice everyday, how does the LLM deal with ethical scenarios like the trolley problem for example? This is a fascinating topic and one we spend a lot of time thinking about. In this episode Dev and Doc, Zeljko Kraljevic and I cover all the up to date topics around reinforcement learning, the benefits and where it can go wrong. We also discuss different RL methods including the algorithms used to train ChatGPT (RLHF). Dev and Doc is a Podcast where developers and doctors join forces to deep dive into AI in healthcare. Together, we can build models that matter. 👨🏻‍⚕️Doc - Dr. Joshua Au Yeung - https://www.linkedin.com/in/dr-joshua... 🤖Dev - Zeljko Kraljevic https://twitter.com/zeljkokr The podcast 🎙️ 🔊Spotify: https://open.spotify.com/show/3QO5Lr3... 📙Substack: https://aiforhealthcare.substack.com/ Hey! If you are enjoying our conversations, reach out, share your thoughts and journey with us. Don't forget to subscribe whilst you're here :) 🎞️ Editor- Dragan Kraljević https://www.instagram.com/dragan_kral... 🎨Brand design and art direction - Ana Grigorovici https://www.behance.net/anagrigorovic...00:00 Highlights 01:27 start 4:38 aligning ethics of ai models 7:04 doctors ethical choices daily 8:00 RLHF and AI training methods 16:29 reinforcement learning 19:35 Preference model -rewarding models correctly can make or break the success 27:05 exploiting reward function, model degradation (and how to fix it) Ref AI intro paper - https://pn.bmj.com/content/23/6/476 Open AI RLHF paper - https://arxiv.org/abs/1909.08593 War and peace of LLMs! - https://arxiv.org/abs/2311.17227