Training AI to act more like a scientist

Teaching a child to solve puzzles offers two paths: one where the child explores through trial and error, and another where they are guided by basic principles and tips. Similarly, the incorporation of rules, such as the laws of physics, into AI training can make these systems more efficient and reflective of the real world. However, determining the significance of various rules in AI systems presents a challenge.

Researchers have developed a framework to assess the importance of rules versus data in "informed machine learning models," which blend both elements. This framework aids in enriching AI with fundamental principles of the real world, improving its ability to solve complex mathematical problems and optimise conditions in chemical experiments.

"Embedding human knowledge into AI models has the potential to improve their efficiency and ability to make inferences, but the question is how to balance the influence of data and knowledge," says Hao Xu of Peking University. The developed framework evaluates different rules to enhance the predictive capabilities of deep learning models.

Generative AI models like ChatGPT and Sora rely solely on data, lacking the capacity to learn physical laws or excel in scenarios not covered by their training data. Informed machine learning offers an alternative by providing models with underlying rules to guide their training, although the relative importance of rules versus data in enhancing model accuracy is not well understood.

"We are trying to teach AI models the laws of physics so that they can be more reflective of the real world, which would make them more useful in science and engineering," states Yuntian Chen of the Eastern Institute of Technology, Ningbo. The researchers developed a method to quantify the contribution of individual rules to a model's accuracy, allowing for model optimisation by adjusting the influence of various rules and removing those that are redundant or conflicting.

This approach has broad applications in fields such as engineering, physics, and chemistry. By optimising machine learning models for solving multivariate equations and predicting the outcomes of thin-layer chromatography experiments, the researchers demonstrate the method's potential to improve experimental conditions in chemistry.

Future plans include creating a plugin tool for AI developers and training models to extract knowledge and rules directly from data. "We want to make it a closed loop by making the model into a real AI scientist," Chen envisions.

This research was supported by institutions including the National Centre for Applied Mathematics Shenzhen and the National Natural Science Foundation of China, marking a significant advancement in the integration of real-world knowledge and rules into AI systems.