For ChatGPT, human-feedback's goal is to fix the wrong data in policy-model's dataset.
There is no essential difference between reinforce learning and supervised learning, here.
Is it right?
Christian Schmidt
Tong Guo There are quite a few differences between reinforcement learning from human feedback (RLHF) and supervised learning. Here are some ressources which adress this in detail:
https://www.marktechpost.com/2023/03/05/5-reasons-why-large-language-models-llms-like-chatgpt-use-reinforcement-learning-instead-of-supervised-learning-for-finetuning/#:~:text=Supervised%20learning%20can%20be%20used,RLHF%20performs%20better%20than%20SL.
https://www.reddit.com/r/reinforcementlearning/comments/zqfw7r/why_cant_we_do_supervised_learning_in_step_3_of/
1 votes
1 thanks
- Badges
- Science topic
- Similar topics
- Biological Science
- Histology
- Fixatives
More Tong Guo's questions See All
Similar questions and discussions