A University of Arizona astronomer has developed a revolutionary method that could fundamentally change the way artificial intelligence models are trained and deployed in science and industry, addressing one of the most critical problems in AI: models that confidently provide incorrect answers.
Space and AI – illustrative photo. Image credit: Pixabay (Free Pixabay license)
Peter Behroozi, associate professor at Stewards Observatorycreated a technique that helps AI systems recognize when their predictions might be unreliable – including models with billions or even trillions of parameters like those powering today’s AI applications. Behroozi paperwhich is awaiting peer review, is available on the open access site arXiv, the free online archive of scientific preprints.
Development of the technique was supported by an exploratory research grant from the National Science Foundation, which funds high-risk, high-reward research. Now that Behroozi’s paper has been published on arXiv, the accompanying code is available for public use, allowing researchers around the world to apply the method to their own work.
The method adapts ray tracing – the computer graphics technique used to create realistic lighting in animated films – to explore the complex mathematical spaces in which AI models operate.
“Current AI models suffer from erroneous but confident results,” Behroozi said. “There are many examples of neural networks ‘mind-blowing’ or inventing non-existent facts, research papers and books to support their incorrect conclusions. This leads to real human suffering, including incorrect medical diagnoses, denied rental applications and facial recognition gone wrong.”
Behroozi’s path to this breakthrough began with his own research into galaxy formation. As the creator of the Universe Machine – a computational framework designed to understand how galaxies form by absorbing large amounts of telescope data – he faced a persistent challenge: existing methods for exploring uncertainty in complex models were simply not suited to the scale and complexity of modern data.
“Galaxies are very complex and so they potentially have many parameters that can control what they do,” Behroozi explained. “For me, existing methods didn’t work very well for exploring the behavior of these parameters.”
The solution came from an unexpected source: a computational physics assignment on how light travels through Earth’s atmosphere, presented during office hours by a University of Alberta undergraduate student. This problem – which required simulating how the speed of light changes as it travels through the atmosphere – gave rise to the idea of ray tracing, the same technique used to create animated films by studios such as Pixar.
“Instead of doing this in three dimensions, I figured out how to make it work in a billion dimensions,” Behroozi said.
The new method applies a gold standard technique called Bayesian sampling, which has been used on small models for decades but is too computationally expensive to apply to modern neural networks. Rather than relying on the prediction of a single model, Bayesian sampling trains thousands of different models on the same data using a special mathematical approach that allows them to explore the diversity of possible answers.
“What happens is, instead of consulting just one expert, you consult all the experts,” Behroozi explained. “If it’s something the experts haven’t seen before, then you’ll get a whole range of answers. And you might conclude that you shouldn’t trust the results that come out.”
Behroozi’s method is several times faster than previous approaches and could lead to safer, more resilient neural networks with far fewer hallucinations. The implications extend far beyond astronomy. Individuals and organizations are increasingly using AI in critical decision-making in medicine, finance, housing, energy, criminal justice and autonomous vehicles. Behroozi’s method would give these systems the ability to recognize when they are uncertain – essentially knowing when they don’t.
“Suppose a doctor orders a routine exam and decides that you need to start cancer treatment immediately, even if you have no other symptoms,” Behroozi said. “Many people in this situation would seek a second opinion. The new method would have a similar effect: instead of the opinion of an AI doctor, it would give a range of plausible opinions.”
For scientists, the method addresses a pervasive problem that undermines trust in AI-assisted research. AI models are used to design new drugs and materials, predict the weather, produce images of black holes, summarize scientific papers and write software – but wrong but confident answers remain all too common.
“This undermines public confidence in scientific results such as weather forecasts and leads scientists to be hesitant to accept new findings based on AI models without separate and costly validation,” Behroozi wrote in his research summary.
For his own work, the technique opens up remarkable new possibilities. Rather than creating simulations that simply match the statistical properties of the universe, Behroozi will be able to determine the actual initial conditions of our universe – essentially creating a movie about the real history of the formation of cosmic structure.
“In the past, we just made galaxies in a universe that looks nothing like ours,” he explained. “This technique allows us to determine what the initial conditions of the real universe were.”
Source: University of Arizona