Navy Sonar AI Spots Hidden Objects With 99.7% Less Training

Finding things underwater is hard. Sonar works like a camera that paints pictures with sound instead of light, and the images come back grainy, echo-heavy, and full of lookalikes — a rock can resemble a mine. Navies rely on AI to help, but there are very few real examples of targets to learn from, and a human often has to double-check every call. That makes training good detection software slow and expensive.

The new approach sidesteps the data problem. The team started with DINOv3, a vision AI already trained on millions of ordinary photographs, and used a technique called LoRA (tuning a small number of settings while leaving the rest of the model frozen). Think of it as adjusting a few knobs rather than rebuilding the machine. This let the model shift from daylight photos to underwater acoustics while training only 0.26% of its weights. The payoff was large: its ranking accuracy jumped from 0.30 to 0.68, meaning real targets surfaced far above false alarms.

Then came the surprise. The researchers added two follow-up stages — one that deliberately fed the AI confusing rocks and sediment, another that pushed targets and clutter further apart. Neither helped. Both landed within noise of doing nothing. Their conclusion: the cheap first step had already done the heavy lifting, and the extra layers were polishing data the model had effectively already learned.

The practical lesson stretches well beyond submarines. If a general-purpose AI can pick up a specialist skill with a sliver of the usual training, then fields short on data — medical scans, factory inspection, wildlife monitoring, search-and-rescue drones — could adopt AI faster and cheaper. The caveats are real: this is military-funded research on one at-sea dataset, and it improves detection, not full autonomy. Humans still stay in the loop.