University of California San Diego engineers have developed a method to make large language models learn new tasks using significantly less data and computing power, reducing the number of parameters required by up to 408 times.
Large language models are made up of billions of parameters that determine how they process information. Traditional fine-tuning methods adjust all of these parameters, which can be costly and prone to overfitting, when a model memorises patterns instead of truly understanding them, causing it to perform poorly on new examples.
The new method updates only the parts that matter most instead of retraining an entire model from scratch. As a result, the approach cuts costs and is more flexible and better at generalising what it learns compared to existing fine-tuning methods.
Higher accuracy with fewer parameters
The researchers showed their method can fine-tune protein language models, which are used to study and predict the properties of proteins, even when very little training data is available. In predicting whether certain peptides can cross the blood-brain barrier, the new method achieved higher accuracy than conventional methods while using 326 times fewer parameters. In predicting protein thermostability, it matched the performance of full fine-tuning while using 408 times fewer parameters.
“With our method, even small labs and startups without huge budgets, supercomputer-level resources or large datasets can adapt large AI models for their own needs,” said Pengtao Xie, a professor in the Department of Electrical and Computer Engineering at the UC San Diego Jacobs School of Engineering. “This work represents a step toward democratising AI.”
The method for fine-tuning and adapting large language models was published in Transactions on Machine Learning Research. The research was supported by the National Science Foundation and National Institutes of Health.
