Dipendra Misra, a researcher at Microsoft Research Labs, explained how LASER can improve the accuracy of large language models at the company's research forum in January.
Using selective class reduction techniques, researchers can intervene and replace matrices with large weights with matrices with small weights.
Weights are an important element that plays a crucial role in the learning and prediction capabilities of artificial neural networks.
Weights in artificial neural networks are similar to synapses in biological neural networks.
The more the linguistic model relies on weights, the greater the weights. According to Microsoft tests, replacing high-weight arrays with low-weight arrays will not affect the accuracy of large language models.
“One might expect increased model loss in interventions with selective class reduction in LLM, which means that LLM performance decreases when we reduce the information of LML trained on large datasets,” explains Misra.
"We found that if we intervene correctly by reducing the size of the selective category, the loss of large language models does not increase, but rather decreases," he added.
The Microsoft team has successfully used selective class reduction in three different major open source language models: RoBERTa, Llama 2, and GPT-J.
Improvements in large language models are sometimes as high as 30 percentage points. The performance of GPT-J, a large, open-source language model for predicting sex based on life history, increased from 70.9% accuracy to 97.5% accuracy after the intervention using a selective class measure of reduction.
AI models make many mistakes in the real world, so the accuracy of large language models remains an issue.
It's not about being afraid of hallucinations, it's not about doing something wrong, it's about inventing something.
Hallucinations and inaccurate AI models can cause significant harm.