AI-LA: Aphasia in Artificial Intelligence Large Language Models

Most AI research focuses on adding capabilities to LLMs. In contrast, little has been done on how to remove these capabilities from pre-trained LLMs.

Finding an approach that scores well on specificity and generalization

A model editing technique scores well on specificity if related facts do not change after the model is edited. A technique scores well on generalization if the fact change is robust to adding or changing the context.

There are three types of approaches to updating parameters - fine-tuning, hyper-networks, and causal tracking.

In this model, I will test all three types!

MODEL GOAL

Reproduce

• Fine-tuning
• Hyper-networks
• causal tracking

Talk to me: https://www.linkedin.com/in/alessandra-faria-b0816053/