A Fruit-Fly Brain Inspired an AI That Learns New Smells Without Forgetting

A fruit fly can separate useful scent signals with a brain smaller than a poppy seed. Researchers at the Okinawa Institute of Science and Technology have now copied part of that strategy in an algorithm that learns odor categories from few examples and adds new ones without readily erasing the old.

KEY NUMBER
A hidden layer of 1,000 simulated neurons turns each incoming scent into a sparse activity pattern that acts like a barcode.

How the fly-inspired shortcut works

Odor readings enter a three-layer spiking neural network. Random connections expand each signal into the 1,000-neuron middle layer, but only a small fraction of those neurons fire for any one smell. An associative learning rule then connects that sparse pattern to an output label.

This arrangement is inspired by the fruit fly's olfactory circuit, where most candidate neurons are suppressed and a small active group represents each odor. The researchers tested Spi-Fly against several classifiers using two scent databases and reported its strongest relative performance when only a few labelled examples were available.

Learning a new smell without wiping the old list

Conventional models can suffer catastrophic forgetting, where training on new classes damages performance on classes learned earlier. Sparse patterns reduce how much one odor's representation overlaps another, allowing Spi-Fly to add categories with less interference.

That feature matters for a portable detector because field data rarely arrive as one complete, balanced training set. A food-storage monitor or pollution sensor may encounter a new compound after deployment and need to update locally.

Why scent is a difficult test

Real air contains mixtures, shifting concentrations, humidity and background chemicals. Biology also shows that odor responses can be highly specific: NewTqnia recently reported that three mosquito species preferred different human scent patterns.

Spi-Fly was evaluated on prepared datasets, not on uncontrolled mixtures. Its overall accuracy still needs improvement against the best conventional methods, according to the researchers.

What happens next

The OIST team plans to connect Spi-Fly to odor-sensing hardware developed with researchers at Eindhoven University of Technology and Kiel University. A convincing next test would measure accuracy, learning speed and actual energy use on mixed airborne chemicals outside a curated dataset.

Takeaway

Spi-Fly offers a compact way to learn new scent labels without rebuilding a model from scratch. Its practical value will depend on whether the same advantage survives contact with a real sensor, mixed odors and measured hardware power limits.