Experimental fruit fly-inspired algorithm classifies odors from few samples

A brain smaller than a poppy seed contains the blueprint for an algorithm that learns smells from very little data. At the Okinawa Institute of Science and Technology, researchers developed Spi-Fly, a fruit fly-inspired system that performed best among the tested methods for few-shot odor classification—recognizing scents after seeing only a few samples.

The model borrows the fly’s way of separating signals. Input odor signals are sent through sparse, random connections to a hidden middle layer of 1,000 neurons. Only a few activate for each scent, producing a distinctive pattern that works like a barcode. Output neurons then associate that pattern with an odor label. When a new scent arrives, the learning rule can add it without wiping out earlier associations.

The team tested Spi-Fly on two scent databases and compared it with a range of other classification algorithms. The researchers also report good continual-learning performance: the system learned new odors while retaining its previous classifications. That addresses catastrophic forgetting, a problem in which standard training methods lose older knowledge as they learn new material.

In practical terms, the approach could make scent-classification systems more useful where data and energy are limited, including explosive detection, allergen assessment, food safety and drug identification. The algorithm was designed for compatibility with neuromorphic hardware—chips built around brain-inspired processing—and researchers at TU Eindhoven and Kiel University are developing odor-sensing hardware that they plan to integrate with Spi-Fly.

The result is still a laboratory-stage algorithm, not a deployed artificial nose. Spi-Fly’s overall performance could be improved compared with the best traditional machine-learning methods, and the team has not yet established how well it handles mixed signals from background smells. The researchers say the same architecture could theoretically extend beyond odors, but that broader use remains untested.