UC San Diego Engineers Develop AI Fine-Tuning Method Requiring Significantly Less Data
Engineers at the University of California San Diego have unveiled a new method for training large language models (LLMs) that dramatically reduces the amount of data and computing power needed, potentially accelerating advancements across numerous AI applications.
Traditional methods of refining LLMs – the technology behind chatbots and tools for complex tasks like protein sequencing – adjust billions of parameters within the model, a process that is both expensive and can lead to “overfitting,” where the model memorizes data instead of generalizing knowledge. The UC San Diego team’s approach focuses on updating only the most critical parameters, resulting in a more efficient and flexible system. This breakthrough could lower the barrier to entry for researchers and smaller organizations seeking to leverage the power of LLMs.
Researchers demonstrated the method’s effectiveness with protein language models, used to study and predict protein properties. In one test predicting whether peptides can cross the blood-brain barrier, the new method achieved higher accuracy than conventional methods while utilizing 326 times fewer parameters. Similarly, in predicting protein thermostability, it matched the performance of full fine-tuning using 408 times fewer parameters. For more information on the fundamentals of large language models, see IBM’s overview of LLMs.
“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 democratizing AI.” The research, published in Transactions on Machine Learning Research, was supported by the National Science Foundation and National Institutes of Health.
The team plans to further refine the method and explore its application to a wider range of LLM tasks, potentially unlocking new possibilities in fields like drug discovery and materials science.
Article Content
Engineers at the University of California San Diego have created a new method to make large language models (LLMs) — such as the ones that power chatbots and protein sequencing tools — learn new tasks using significantly less data and computing power.
LLMs 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 memorizes patterns instead of truly understanding them, causing it to perform poorly on new examples.
The new method developed by UC San Diego engineers takes a smarter approach. Instead of retraining an entire model from scratch, it updates only the parts that matter most. As a result, the new method cuts costs and is more flexible and better at generalizing what it learns compared to existing fine-tuning methods.
The researchers showed that 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 are available. For example, 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 democratizing AI.”
The new method for fine-tuning and adapting LLMs was published in Transactions on Machine Learning Research. This research was supported by the National Science Foundation (IIS2405974 and IIS2339216) and National Institutes of Health (R35GM157217 and R21GM154171).
Learn more about research and education at UC San Diego in:
