ALBUQUERQUE, N.M. — At a time when artificial intelligence is increasingly relied upon to analyze everything from satellite imagery to national security data, researchers at Sandia National Laboratories are tackling one of the field’s most stubborn problems: how to teach machines something new without causing them to forget what they already know.
Leading the effort is Tian Ma, who is overseeing a three-year research project known as Adaptive Intelligence for Global Monitoring Lite, or AIGM Lite. The project explores a new approach to artificial intelligence that could allow analytical systems to update themselves more effectively as new information becomes available.
Tian Ma
The challenge is one that extends far beyond computer science laboratories.
Most machine learning systems are updated through a lengthy process of collecting new data, retraining models, and redeploying them. By the time that cycle is complete, real-world conditions may have already changed. For systems used in rapidly evolving environments, that lag can reduce reliability and limit effectiveness.
Ma’s team is studying whether AI can learn continuously rather than waiting for periodic overhauls.
The approach borrows from Nobel Prize-winning psychologist Daniel Kahneman’s concept of “Thinking, Fast and Slow.” Researchers have created a dual-system model in which one component provides rapid analysis while a second, slower process evaluates new information, monitors performance and prepares updates over time.
The goal is to make artificial intelligence both adaptable and dependable.
A major obstacle in the field is what’s known as catastrophic forgetting. When many AI systems learn new information, they often lose accuracy when performing tasks they previously handled well.
To address that problem, the Sandia team designed a system that allows models to absorb new information while preserving prior knowledge.
Early testing has produced promising results.
In controlled trials simulating the gradual arrival of new data, traditional incremental-learning methods achieved roughly 70% overall accuracy. The AIGM Lite approach improved performance to approximately 90% accuracy, according to researchers.
The technology is being developed primarily for sensor-based analytical systems rather than conversational AI platforms. Potential applications include space-based monitoring, national security analysis and other missions that depend on continuous streams of data from satellites and remote sensors.
As new objects, patterns and conditions emerge, the system is designed to recognize information that falls outside its existing knowledge base and flag those cases for further review rather than automatically making assumptions.
Researchers say that capability could become increasingly important as modern monitoring systems generate larger volumes of data than ever before.
While the project remains in development, Sandia scientists believe the work could help create AI systems that remain effective as conditions evolve, without sacrificing the knowledge they’ve already gained.
“Mission-ready AI must do more than classify the known,” Ma said. “It must adapt to emerging conditions without losing prior knowledge.”
