Microsoft LASERs away LLM inaccuracies
Illustration: The Verge
During the January Microsoft Research Forum, Dipendra Misra, a senior researcher at Microsoft Research Lab NYC and AI Frontiers, explained how Layer-Selective Rank Reduction (or LASER) can make large language models more accurate.
With LASER, researchers can “intervene” and replace one weight matrix with an approximate smaller one. Weights are the contextual connections models make. The heavier the weight, the more the model relies on it. So, does replacing something with more correlations and contexts make the model less accurate? Based on their test results, the answer, surprisingly, is no.
“We are doing intervention using LASER on the LLM, so one would expect that the model loss should go up as we are doing more approximation, meaning that the model is going to perform bad, right, because we are throwing out information from an LLM, which is trained on large amounts of data,” Misra said. “But to our surprise, we find that if the right type of LASER intervention is performed, the model loss doesn’t go up but actually goes down.”
Misra said his team successfully used LASER on three different open-source models: RoBERTa, Llama 2, and Eleuther’s GPT-J. He said, at times, model improvement increased by 20 to 30 percentage points. For example, the performance of GPT-J for gender prediction based on biographies went from 70.9 percent accuracy to 97.5 percent after a LASER intervention.
AI models make a lot of factual mistakes, so LLM accuracy remains a concern, and it’s not just fear of hallucinations, which are less about getting things wrong and more about making things up. Hallucinations and inaccurate AI models can be entertaining, but they can do considerable harm, too.
Illustration: The Verge
During the January Microsoft Research Forum, Dipendra Misra, a senior researcher at Microsoft Research Lab NYC and AI Frontiers, explained how Layer-Selective Rank Reduction (or LASER) can make large language models more accurate.
With LASER, researchers can “intervene” and replace one weight matrix with an approximate smaller one. Weights are the contextual connections models make. The heavier the weight, the more the model relies on it. So, does replacing something with more correlations and contexts make the model less accurate? Based on their test results, the answer, surprisingly, is no.
“We are doing intervention using LASER on the LLM, so one would expect that the model loss should go up as we are doing more approximation, meaning that the model is going to perform bad, right, because we are throwing out information from an LLM, which is trained on large amounts of data,” Misra said. “But to our surprise, we find that if the right type of LASER intervention is performed, the model loss doesn’t go up but actually goes down.”
Misra said his team successfully used LASER on three different open-source models: RoBERTa, Llama 2, and Eleuther’s GPT-J. He said, at times, model improvement increased by 20 to 30 percentage points. For example, the performance of GPT-J for gender prediction based on biographies went from 70.9 percent accuracy to 97.5 percent after a LASER intervention.
AI models make a lot of factual mistakes, so LLM accuracy remains a concern, and it’s not just fear of hallucinations, which are less about getting things wrong and more about making things up. Hallucinations and inaccurate AI models can be entertaining, but they can do considerable harm, too.