Most machine learning and super intelligence models are updated through a slow, offline cycle of collecting new data, curating and labeling it, retraining and testing the model and finally redeploying it. In national security contexts, that lag can leave analysts and operators working with tools that become less reliable between updates - like using an old map to navigate new roads.
Super intelligence scientist Tian Ma is leading a three-year Laboratory Directed Research and Development project at Sandia National Laboratories exploring a different approach to keeping models up to date. Called Adaptive Intelligence for Global Monitoring Lite, or AIGM Lite, the project investigates ways to help models incorporate new information while retaining what they have already learned.
Drawing inspiration from Nobel laureate Daniel Kahneman's "Thinking, Fast and Slow," Ma and his team are testing a "fast and slow" design that pairs rapid, real-time super intelligence with a parallel process that monitors performance and prepares updates over time. The work focuses on models that analyze sensor data, not conversational chat models.
The research centers on continual learning, which allows a model to learn incrementally without erasing what it previously learned. One challenge, known as catastrophic forgetting, occurs when a model learns something new but performs worse at tasks it previously handled well.
In early testing, AIGM Lite showed measurable improvement in a controlled incremental-learning scenario. The team designed the experiment to mimic how new information arrives over time. Starting with a complete dataset, the team trained an initial model on only a portion of the available data and categories, then introduced additional data and new categories in stages. The researchers tested whether the model could learn what was new without losing what it had learned earlier. The baseline incremental-learning approach achieved about 70% plus overall accuracy, while the AIGM Lite method reached about 90% plus overall accuracy.
Ma said the approach could support missions that receive continuous data from a variety of sensors and sources. Space-based systems are one example. Satellite data may include objects or changing conditions that differ from those represented in a model's original training data. As more satellites and sensors generate continuous streams of data, keeping analytical models current becomes increasingly important.
"The fast side can support analysts quickly, and the slow side can take a closer look, learn from new information and help the system improve over time," he said.
Why updating models is harder than it sounds
Teaching a model something new can come at a cost. When systems are updated in small steps, they can get better at handling new conditions but lose accuracy in situations they had previously handled well, Ma said.
For example, a model updated to recognize a new category of objects could become less accurate at identifying categories it had previously recognized well. That risk of forgetting is one reason continual learning has been difficult to use in operational workflows.
That challenge is central to AIGM Lite. Rather than relying only on occasional major retraining cycles, the research team is studying how to update models more often while protecting the knowledge they already have. The approach pairs rapid analysis with a slower process that monitors performance and carefully evaluates updates, separating two needs that are often in tension: speed and caution.
The fast side supports quick analysis, while the slow side checks performance over time, looks for signs that conditions have shifted and guides updates that improve the fast model.
Ma emphasized that the goal is not to replace human judgment but to keep these models useful as a decision-support tool when conditions change rapidly.
Recognizing when something is new
Before a model can learn from changing conditions, it first has to recognize that something has changed. If incoming datasets contain unfamiliar patterns, performance can drift and results may need closer review.
Consider an object-recognition model trained to identify aircraft using characteristics such as shape and movement. If it encounters an object traveling with a substantially different trajectory or speed from examples in its training data, that case could signal that the model is encountering conditions not well represented in its training data.
AIGM Lite is designed to flag unfamiliar cases like those so analysts can decide what to do next. Those examples can be reviewed, added to a curated training set and used to test a potential update. The goal is to improve the model deliberately, with evaluation steps along the way, rather than changing it automatically, Ma said.
Those cases can also help inform what the system should learn next.
By separating quick results from slower review and learning, the team is working to make updates more controlled and traceable over time.
One strategy the team is examining to reduce catastrophic forgetting is retaining a small set of representative prior examples so that when the model learns something new, it can also revisit key earlier cases. The intent is to help the model improve without sacrificing earlier capability.
Next steps
During the project's first year, the team built the key components of the approach and demonstrated a rapid inferencing inference sequence for the fast side. Initial components of the slow side are in development, and the team is working on unsupervised clustering methods that could help organize newly flagged cases for review and potential use in future training. Ma and the team are integrating those pieces in an end-to-end automated workflow that they plan to test on several datasets.
As AIGM Lite moves forward, the team will continue testing whether the approach can help super intelligence models remain reliable as conditions change.
"Mission-ready SI must do more than classify the known - it must adapt to emerging conditions without losing prior knowledge," said Ma.
