When imparting puzzle-solving skills to a child, you can either allow them to discover the solution through trial and error, or you can offer guidance with a set of fundamental principles and pointers.
Similarly, incorporating laws and principles, such as those of physics, into AI training can enhance efficiency and reflect the real world more accurately. However, determining the value of various rules for AI can be a challenging task.
In the journal Nexus, a team of researchers reveals on March 8 that they have devised a framework for evaluating the relative significance of rules and data in informed machine learning models that combine both elements. They demonstrated that by doing so, they could assist the AI in integrating basic laws of the physical world and navigating scientific problems, such as solving complex mathematical equations and optimizing experimental conditions in chemistry experiments.
Generative AI models like ChatGPT and Sora operate solely on data—
An alternative approach is informed machine learning, where researchers provide the model with some fundamental rules to guide its training process. However, little is known about the relative importance of rules versus data in determining model accuracy.
To enhance the performance of informed machine learning, the team developed a framework for calculating the contribution of a single rule to a model's predictive accuracy. They also examined the interactions between different rules, as most informed machine learning models incorporate multiple rules, and having too many rules can lead to model breakdown.
This enabled them to optimize models by adjusting the influence of various rules and filtering out redundant or conflicting rules. They also identified synergistic rules and others that rely entirely on the presence of other rules.
The researchers say that their framework has practical applications in engineering, physics, and chemistry. In their paper, they showcased the potential of their method by using it to optimize machine learning models for solving multivariate equations and predicting the results of thin layer chromatography experiments, thus improving future experimental chemistry conditions.
Next, the researchers plan to develop their framework into a plugin tool that can be utilized by AI developers. Eventually, they aim to train their models to extract knowledge and rules directly from data, rather than relying on rules selected by human researchers.