AI that learns to generalize: less supervision, greater reach
Most current systems work well within the domain they were trained for. When they encounter something outside that scope—a class they never saw, a different domain, a task for which there are no examples—their performance drops. GPAIs are AI systems capable of generalizing with less labeled data, less supervision, and greater ability to adapt to new contexts.
GPAIs are categorized into two main approaches
In the closed world, the system operates on a known and predefined set of tasks. It can address various tasks, but all are part of the scenario anticipated during its design and training, so its ability to act is tied to that defined framework.
In the open world, the system operates in a dynamic and changing environment, where new tasks, data, or unforeseen situations may emerge. Therefore, beyond solving what is already known, it must demonstrate greater generalization capacity, adaptation, and response to the unknown.
Four research lines that the group works on in parallel, with applications in health, industry, and rural environments
Learning without prior examples through semantic knowledge transfer.
Two complementary lines: Contrastive Learning for biomedical signals and SSL for tabular data in highly complex environments.
Automated machine learning and learning from minimal examples for rapid adaptation.
Autonomous systems with decision-making and strategic planning capabilities.