Telegram Group & Telegram Channel
Telegram Group » United States » Parallel Experiments » Telegram Webview » Post 938
Parallel Experiments
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。
arXiv.org
Direct Preference Optimization: Your Language Model is Secretly a...
While large-scale unsupervised language models (LMs) learn broad world knowledge and some reasoning skills, achieving precise control of their behavior is difficult due to the completely...
www.tg-me.com/us/Parallel Experiments /com.LinghaoCh/938
1.5K viewsLinghao Zhang, edited  



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: |

Telegram Be The Next Best SPAC

I have no inside knowledge of a potential stock listing of the popular anti-Whatsapp messaging app, Telegram. But I know this much, judging by most people I talk to, especially crypto investors, if Telegram ever went public, people would gobble it up. I know I would. I’m waiting for it. So is Sergei Sergienko, who claims he owns $800,000 of Telegram’s pre-initial coin offering (ICO) tokens. “If Telegram does a SPAC IPO, there would be demand for this issue. It would probably outstrip the interest we saw during the ICO. Why? Because as of right now Telegram looks like a liberal application that can accept anyone - right after WhatsApp and others have turn on the censorship,” he says.

The STAR Market, as is implied by the name, is heavily geared toward smaller innovative tech companies, in particular those engaged in strategically important fields, such as biopharmaceuticals, 5G technology, semiconductors, and new energy. The STAR Market currently has 340 listed securities. The STAR Market is seen as important for China’s high-tech and emerging industries, providing a space for smaller companies to raise capital in China. This is especially significant for technology companies that may be viewed with suspicion on overseas stock exchanges.

Parallel Experiments from us


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。

 Parallel Experiments TG
Webview: 938
Parallel Experiments.Telegram Webview
Parallel Experiments Telegram TG Channel
Telegram Updated:
Telegram Parallel Experiments
FROM USA