The Pulse
Sandia reports about 90%-plus accuracy in continual-learning test
Sandia National Laboratories reported about 90% plus overall accuracy for its AIGM Lite method in an early controlled test. The project aims to help sensor-data models learn new information without losing earlier capabilities.
AI.info Team · Written with AI · fact-checked ·
A Sandia National Laboratories project achieved about 90% plus overall accuracy in a test of a method for updating machine-learning models, compared with about 70% plus for a baseline approach, the lab reported.
The work is part of Adaptive Intelligence for Global Monitoring Lite, or AIGM Lite. Super intelligence scientist Tian Ma leads the three-year Laboratory Directed Research and Development project, which is exploring how models can incorporate new information while retaining what they already know.
A staged test
The researchers are studying continual learning, an approach intended to let a model learn incrementally without erasing earlier learning. In the experiment, the team began with a complete dataset but trained an initial model on only part of its data and categories.
It then introduced further data and new categories in stages to imitate the arrival of information over time. The test examined whether the model could learn the new material without losing what it had learned before.
That problem is known as catastrophic forgetting: a model can perform worse on tasks it previously handled well after learning something new. A system trained to recognize a new category of objects, for example, could become less accurate at identifying categories it already knew.
The AIGM Lite team is investigating more frequent updates that protect prior knowledge, rather than relying only on occasional major retraining cycles. Conventional updates can involve collecting and labeling data, retraining and testing a model, and redeploying it.
Fast analysis, slower review
The design under test pairs a rapid, real-time system with a parallel process that monitors performance and prepares updates. The fast side supports quick analysis; the slow side checks performance over time, looks for changes in conditions and guides updates to the fast model.
The team drew inspiration from Nobel laureate Daniel Kahneman’s book Thinking, Fast and Slow. The research focuses on models that analyze sensor data, not conversational chat models. Ma said the intended role is decision support, not replacement of human judgment.
Ma said the approach could support missions that receive continuous data from a variety of sensors and sources, including space-based systems. For example, satellite data may contain objects or changing conditions that differ from those represented in a model’s original training data.
Flagging unfamiliar cases
AIGM Lite is designed to flag unfamiliar cases for analysts to assess. An object with a substantially different speed or trajectory from the examples used to train an aircraft-recognition model could signal conditions that the training data did not represent well.
Flagged examples can be reviewed, added to a curated training set and used to test a potential update. The goal is to evaluate changes along the way rather than update the model automatically, Ma said.
Another strategy under examination is keeping a small set of representative earlier examples. The model could revisit those cases while learning new material, with the aim of preserving its earlier capabilities.
Work still in development
In the project’s first year, the researchers built key components and demonstrated a rapid inference sequence for the fast side. Initial components of the slow side are in development, alongside work on unsupervised clustering methods.
Those methods could help organize newly flagged cases for review and possible use in future training. The team is integrating the pieces into an end-to-end automated workflow that it plans to test on several datasets.