Why chatGPT/LLM should have unlearning capability like human has..

Executive Summary

Do you know, chatGPT/LLM has this open problem to solve. This problem(unlearn) has potential to negate the power of chatGPT.

1. The article starts with Why it is important
2. Then it talks about current state.
3. Article finishes telling about opportunities for innovation and the mitigation suggestion

Why?

Look at below pic. This is corrections published by a renowned news paper. Think, If chatGPT has been trained on 25th data, then it acquired an incorrect knowledge. Can chatGPT forgets wrong info based on below correction advertisement?

Some of these corrections could be quite important . Look at the below correction related to a crime.

So, the bigger question is: How can chatGPT/LLM correct itself?

Additional case

Look at the below example. James Gunn?was fired from "Guardians of the Galaxy 3"?by Disney after his offensive tweets resurface. He was fired in 2018, for tweets that were written between 2008 and 2011.

Look at James comment. He mentioned that he has changed himself and like his new avatar.

“I used to make a lot of offensive jokes,” Gunn wrote. “I don’t anymore. I don’t blame my past self for this, but I like myself more and feel like a more full human being and creator today.?Love you to you all.”

So, the question is that should a person's content be hanging around forever by some LLM?

There is option in facebook to delete your content. What if LLM learns it and keep on using it even though it doesn't exist at source?

Current capability of chatGPT

I asked chatGPT about any instance where it learned wrong info and what was the measure taken. See the answer in the highlight

As a human being, we are expected to unlearn the incorrect knowledge. What about chatGPT?

The simplest approach would be to delete the data point from the training data and retrain the model. However, this is clearly expensive. For example, OpenAI spent an estimated cost of between 10 and 20 million dollars?to train?GPT-3. Thus, we need better and cheaper alternatives.

Mitigation

If chatGPT/LLM is trained cautiously, then it will be less likely to suffer this problem.

1. Layered training of chatGPT. So, chatGPT will be feed with small subset of data at a time. Note that It will help LLM to remove specific layer. Note that this approach also has flaws. Curious chaps can read https://arxiv.org/abs/1912.03817!!
2. The training data needs to be cleaned to avoid any such conditions which triggers unlearn

Opportunities

This challenge opens up window of opportunities as well

1. Innovation: To give a solution which solves this problem for ever.. To start with, I suggest to follow the ongoing research papers/articles
2. Build excellence is training LLM which has lesser and lesser of such issues

Thanks to these helping hands

https://towardsdatascience.com/machine-unlearning-the-duty-of-forgetting-3666e5b9f6e5

https://techcrunch.com/2018/07/20/james-gunn-fired-from-guardians-of-the-galaxy-3-after-offensive-tweets-resurface/

https://techcrunch.com/2018/07/20/james-gunn-fired-from-guardians-of-the-galaxy-3-after-offensive-tweets-resurface/

https://techcrunch.com/2018/07/20/james-gunn-fired-from-guardians-of-the-galaxy-3-after-offensive-tweets-resurface/

https://www.nytimes.com/international/section/corrections

https://www.poynter.org/fact-checking/2017/not-fake-news-just-plain-wrong-top-media-corrections-of-2017/

https://arxiv.org/pdf/1912.03817.pdf

https://arxiv.org/pdf/1912.03817.pdf

https://towardsdatascience.com/machine-unlearning-the-duty-of-forgetting-3666e5b9f6e5

chatGPT platform