Reshaping the Model’s Memory without the Need for Retraining | by Salvatore Raieli | Oct, 2023

#Reshaping #Models #Memory #Retraining #Salvatore #Raieli #Oct


Erasing any echo of problematic content a large language model has learned

Salvatore Raieli
Towards Data Science
large language models machine unlearning
Photo by Drew Saurus on Unsplash

“To forgive is wisdom, to forget is genius. ”
Joyce Cary

Large language models (LLMs) have taken the world by storm. In less than a year they are ubiquitous and are now used by millions of users. These models are often trained with huge amounts of text (including problematic material and sensitive data). How do you make a model forget? The same that could store the entirety of human knowledge?

large language models machine unlearning
Photo by Paul Pastourmatzis on Unsplash

LLMs stand as a testament to both our accomplishments and the challenges that lie ahead — source

LLMs have surprised both users and researchers with their ability to learn from huge amounts of text and identify language patterns and cultural nuances. While they could be the basis for a new application and scientific revolution, they have a dark side.

Huge corpora must be used to train these patterns. While it is true that the greater the amount of data used the better the performance of an LLM, collecting this data is expensive. To limit costs, indiscriminate scraping of data from the Internet is often used. These corpora therefore also contain extremely problematic data: copyrighted texts, toxic or malicious data, inaccurate or fake content, personal data, and more.

large language models machine unlearning
image source: here