A groundbreaking technique developed by a University of Arizona astronomer could revolutionize the way AI models are trained and utilized across various fields, addressing a critical issue in AI: the tendency to produce incorrect answers with high confidence. Peter Behroozi, an associate professor at the Steward Observatory, has created a method that enables AI systems to recognize when their predictions might be unreliable, even for large-scale models with billions to trillions of parameters. The research, available on the open-access arXiv site, was supported by a National Science Foundation grant for high-risk, high-reward research. The technique adapts ray tracing, a computer graphics method for realistic lighting in films, to explore complex mathematical spaces where AI models operate. Behroozi's work began with his research in galaxy formation and the Universe Machine, a computational framework for understanding galaxy formation. He encountered challenges with existing methods for exploring uncertainty in complex models, leading him to develop a novel approach. The solution came from a computational physics homework problem, which inspired him to apply ray tracing to a billion dimensions, a significant advancement from the three-dimensional approach. The new method employs Bayesian sampling, a gold standard technique, to train thousands of models on the same data, exploring diverse responses. This approach is computationally efficient and allows AI systems to recognize when their predictions are uncertain, providing a range of plausible opinions. The implications are far-reaching, impacting critical decision-making in medicine, finance, housing, energy, criminal justice, and autonomous vehicles. By enabling AI to acknowledge its uncertainties, the method enhances trust in AI-assisted research, addressing a pervasive problem that undermines public trust in scientific output. For Behroozi's research, this technique opens up new possibilities, allowing him to determine the initial conditions of the universe and create a movie of its real history. This breakthrough has the potential to significantly improve the reliability and safety of AI models, making them more trustworthy in various applications.
