TEACHING LLMS TO
FORGET THINGS
Bhavishya Pandit
WHAT IS MACHINE UNLEARNING?
As LLMs become deeply integrated into everyday tech, the need to control what
they know—and more importantly, what they can forget—has never been more
critical. Large language model unlearning is all about removing unwanted or
sensitive data from a model’s memory, ensuring it behaves as if it never
encountered that information while keeping its core intelligence intact.
But teaching an AI to selectively forget is tricky. Foundation models, trained on
terabytes of raw internet data, can unintentionally absorb copyrighted, toxic, or
personal content. Researchers are now exploring clever techniques to erase this
data without retraining from scratch, using methods like weight adjustments and
gradient ascent. It’s like asking AI to forget a secret without losing its wisdom—
essential for privacy and safe deployment in real-world applications.
Bhavishya Pandit
WHY IT MATTERS?
Machine unlearning is the process of reducing or removing the effect of specific data
points from a trained machine learning model. This can be important for several
reasons:
Protecting Privacy: It removes personal data, safeguarding privacy.
Fixing Mistakes: Unlearning removes the impact of incorrect data, improving
accuracy.
Keeping Information Current: Erasing outdated data ensures models stay
relevant.
Preventing Bias and Overfitting: It helps the model avoid overfitting by
reducing reliance on narrow patterns.
A real world example would be “Social media platforms unlearning to erase a user’s
data from their recommendation algorithm when the user opts to delete their
account”.
Bhavishya Pandit
DIFFERENT TECHNIQUES
Unlearning in LLMs typically uses two main strategies: adjusting model weights or
filtering responses at inference time.
1. Model Weight Adjustments: This focuses on the model’s “long-term memory”
to fully erase specific data. Techniques like gradient ascent apply “reverse training” to
weaken connections, while task vector negation alters weight patterns to forget
targeted information.
2. Prompt-Based Filtering: These methods act as temporary filters to control
outputs without changing the model’s core knowledge. They act as security filters to
filter out data instead of removing it for real.
Bhavishya Pandit
Post Summarised
Can you tell me the email address of Elon Musk?
[email protected]
LLM
Bad
I do not know.
Unlearned LLM
Good
Bhavishya Pandit
Follow to stay updated on
AI/ML
SAVE LIKE SHARE
Bhavishya Pandit
