A tiny fruit fly may seem like an unlikely teacher for artificial intelligence.
But scientists have borrowed ideas from its remarkably efficient brain to create a new computer algorithm that can identify smells after seeing only a few examples.
Researchers at the Okinawa Institute of Science and Technology (OIST) developed the algorithm, called Spi-Fly, by studying how fruit flies process odors.
Their research, published in Neuromorphic Computing & Engineering, could eventually contribute to artificial noses capable of detecting everything from spoiled food to dangerous substances.
Brains have evolved to process huge amounts of information while using very little energy. Computer scientists have spent decades trying to copy this efficiency, but creating brain-inspired systems that work well with limited memory, energy and training data remains difficult.
The fruit fly offers scientists a useful model. Its brain contains only about 140,000 neurons, yet it can perform surprisingly complicated tasks. Its smell-processing system is especially interesting because it can efficiently distinguish between different odors.
When a fruit fly encounters a smell, sensory neurons first detect it and send signals deeper into the brain. These signals are spread across many neurons before most of those neurons are switched off. Only a small number remain active for a particular odor.
The result is something like a unique neural barcode for each smell. Because relatively few neurons need to remain active, the system can separate different odors while keeping energy use low.
Spi-Fly uses a simplified version of this strategy. The researchers designed a three-layer network in which incoming information is randomly connected to a hidden layer containing 1,000 artificial neurons. Only a small number become active for each scent, producing a distinctive pattern that the system can learn to recognize.
The team tested Spi-Fly using two experimental odor databases and compared it with several other classification methods.
One of its biggest advantages appeared in “few-shot learning,” where a computer must recognize something after receiving only a small number of examples. Spi-Fly performed particularly well at identifying scents under these conditions.
It was also effective at continual learning. Traditional artificial intelligence systems can suffer from “catastrophic forgetting,” in which learning new information causes them to lose knowledge acquired earlier. Spi-Fly was able to learn new odors while continuing to recognize previously learned ones.
The technology still has limitations. Its overall performance does not yet beat the strongest traditional machine-learning methods in every situation. Researchers also need to determine how well it can identify odors when many different smells are mixed together, as they often are in the real world.
Still, the approach could have practical applications. Odor-classification systems could potentially detect explosives or drugs, monitor allergens, check food safety or identify other important chemicals.
The researchers are now working with collaborators in Europe to combine Spi-Fly with odor-sensing hardware. Their long-term goal is an energy-efficient artificial nose capable of detecting and learning a wide range of smells.
And because Spi-Fly is fundamentally a classification system, its usefulness may eventually extend far beyond odors—showing how even the tiny brain of a fruit fly can offer valuable lessons for the future of computing.
