Misra gave an example of the performance of GPT-J for gender prediction based on biographies, which improved from 70.9 percent accuracy to 97.5 percent after a LASER intervention. Despite these improvements, Misra acknowledged that AI models often make factual errors, and their accuracy remains a concern. He also highlighted the potential harm caused by "hallucinations", where AI models make things up, emphasizing the need for continued improvements in model accuracy.
Key takeaways:
- Dipendra Misra, a senior researcher at Microsoft Research Lab NYC and AI Frontiers, discussed the use of Layer-Selective Rank Reduction (LASER) to improve the accuracy of large language models during the January Microsoft Research Forum.
- LASER allows researchers to replace one weight matrix with a smaller one, which surprisingly does not reduce the model's accuracy but can actually improve it.
- Misra's team successfully used LASER on three different open-source models: RoBERTa, Llama 2, and Eleuther’s GPT-J, with model improvement increasing by 20 to 30 percentage points in some cases.
- Despite these improvements, the accuracy of AI models remains a concern due to the potential for factual errors and 'hallucinations', where the model makes things up.