Sound moves through the world in ways our ears cannot fully grasp. Ultrasound — far beyond the threshold of human hearing — has been mastered by nature long before we thought to harness it. Bats navigate the dark with extraordinary precision, dolphins map the ocean floor, and whales communicate across vast distances, all through the sophisticated use of ultrasonic signals.

At L-MAP, we take these biological systems as our starting point. By studying how nature encodes, transmits, and interprets ultrasonic signals, we design sensing and signal processing frameworks that bring the same elegance and efficiency to machines.

Every child learns to recognize a dog without being handed a textbook. They observe, compare, and generalize — building understanding from the world around them, not from explicit instruction. Yet for decades, machines have required exactly that: vast collections of carefully labeled examples, painstakingly annotated by human hands.

At L-MAP, we take inspiration from the way intelligence naturally develops. We explore learning frameworks that move beyond the need for labels — from semi-supervised and zero-shot learning, to knowledge distillation and federated learning. We also investigate how the design of loss functions fundamentally shapes the way a model draws boundaries between categories. Together, these approaches bring us closer to machines that learn the way living things do: flexibly, efficiently, and from the richness of the world itself.

A doctor does not diagnose from a single sense alone. They listen to a patient's history, examine what they see, and draw on years of accumulated knowledge to reach a conclusion. True understanding, in medicine as in life, emerges from the integration of multiple streams of information.

At L-MAP, we build AI systems that do the same. By bringing together audio encoders, vision encoders, and language models, we develop multimodal frameworks capable of reasoning across sound, image, and text — not as separate inputs, but as a unified whole. Our current work focuses on the dental domain, where we are developing a vision-language model trained to interpret clinical imagery and acoustic signals with the depth and nuance that specialist knowledge demands.

This is not simply a technical exercise. It is a step toward AI that does not merely process information, but genuinely understands it.