Telegram Group & Telegram Channel
https://arxiv.org/abs/2305.18290 #llm #ai

今天深入学习了 DPO,再次感叹扎实的数学功底对 AI/ML Research 的重要性……

原始的 RLHF 是用 pairwise human preference data(A 和 B 哪个更好)去训练一个 reward model,然后用 RL 来训练主 policy model,objective 是 minimize negative log likelihood + regularization(比如 PPO 就是通过新旧 policy 之间的 KL Divergence 来做 regularization)。这样的缺点在于 RL 是出了名的难搞,而且还需要一个 critic model 来预测 reward,使得整个系统的复杂性很高。

DPO 的思路是,观察到 RLHF 的 objective 本质上是 minimize loss over (latent) reward function,通过一番 reparameterization 等数学推导,重新设计了一个 minimize loss over policy 的 objective,绕过了中间这个 reward model,让 gradient update 直接增加 policy model 生成 winner response 的概率并降低 loser response 的概率,大幅简化了流程。

拓展阅读:
- KTO: 更进一步,不需要 pairwise comparison,只用对 individual example 的 upvote/downvote 也可以学习到 preference。
- IPO: 解决 DPO 容易 overfit 的问题。



tg-me.com/LinghaoCh/938
Create:
Last Update:

https://arxiv.org/abs/2305.18290 #llm #ai

今天深入学习了 DPO,再次感叹扎实的数学功底对 AI/ML Research 的重要性……

原始的 RLHF 是用 pairwise human preference data(A 和 B 哪个更好)去训练一个 reward model,然后用 RL 来训练主 policy model,objective 是 minimize negative log likelihood + regularization(比如 PPO 就是通过新旧 policy 之间的 KL Divergence 来做 regularization)。这样的缺点在于 RL 是出了名的难搞,而且还需要一个 critic model 来预测 reward,使得整个系统的复杂性很高。

DPO 的思路是,观察到 RLHF 的 objective 本质上是 minimize loss over (latent) reward function,通过一番 reparameterization 等数学推导,重新设计了一个 minimize loss over policy 的 objective,绕过了中间这个 reward model,让 gradient update 直接增加 policy model 生成 winner response 的概率并降低 loser response 的概率,大幅简化了流程。

拓展阅读:
- KTO: 更进一步,不需要 pairwise comparison,只用对 individual example 的 upvote/downvote 也可以学习到 preference。
- IPO: 解决 DPO 容易 overfit 的问题。

BY Parallel Experiments




Share with your friend now:
tg-me.com/LinghaoCh/938

View MORE
Open in Telegram


Parallel Experiments Telegram | DID YOU KNOW?

Date: |

How Does Bitcoin Mining Work?

Bitcoin mining is the process of adding new transactions to the Bitcoin blockchain. It’s a tough job. People who choose to mine Bitcoin use a process called proof of work, deploying computers in a race to solve mathematical puzzles that verify transactions.To entice miners to keep racing to solve the puzzles and support the overall system, the Bitcoin code rewards miners with new Bitcoins. “This is how new coins are created” and new transactions are added to the blockchain, says Okoro.

The global forecast for the Asian markets is murky following recent volatility, with crude oil prices providing support in what has been an otherwise tough month. The European markets were down and the U.S. bourses were mixed and flat and the Asian markets figure to split the difference.The TSE finished modestly lower on Friday following losses from the financial shares and property stocks.For the day, the index sank 15.09 points or 0.49 percent to finish at 3,061.35 after trading between 3,057.84 and 3,089.78. Volume was 1.39 billion shares worth 1.30 billion Singapore dollars. There were 285 decliners and 184 gainers.

Parallel Experiments from ms


Telegram Parallel Experiments
FROM USA