Originally published in Xplore TechnologyMarch 8, 2024.
When teaching a child to solve puzzles, you can either let them solve through trial and error or guide them with some basic rules and tips. Likewise, incorporating rules and tricks into AI training, such as the laws of physics, could make it more effective and better reflect the real world. However, helping the AI evaluate the value of different rules can be a tricky task.
Researchers report March 8 in the newspaper Link that they developed a frame to evaluate the relative value of rules and data in machine learning models” that integrate both. They showed that by doing so, they could help AI integrate fundamental real-world laws and better solve scientific problems such as solving complex math problems and optimizing experimental conditions in chemistry experiments.
“Integrating 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,” explains the first author Hao Xu of Peking University. “Our framework can be used to evaluate different knowledge and rules to improve the predictive ability of deep learning models.”
Generative AI models like ChatGPT and Sora are purely data-driven: the models receive training data and learn themselves through trial and error. However, with only data to work with, these systems have no way of learning physical laws, such as gravity or gravity. fluid dynamicsand they also have difficulty performing in situations that differ from their training data.
To continue reading this article, Click here.
