Researchers Kevin Max and Yang Shen of the Okinawa Institute of Science and Technology have developed an algorithm called Spi-Fly, inspired by the olfactory system of fruit flies. According to the study published in the journal Neuromorphic Computing and Engineering, the algorithm aims to mitigate two main problems with existing “electronic noses”: their need for large numbers of labeled samples to learn new odors and their tendency to forget previously learned odors when learning a new one. This phenomenon is known as “catastrophic forgetting.”
Although electronic noses are used in food quality control, environmental monitoring and security screening, they are generally trained for specific odors. Reconfiguring a system for a different task often requires the software to be retrained almost from scratch. The researchers explain how fruit flies can distinguish and remember odors with their small brains using a method called “sparse coding.” The flies’ approximately 2,000 Kenyon cells process signals from odor receptors randomly and sparsely. The pattern formed by the cells that remain active after receiving the strong inhibitory signals of APL neurons serves as a “barcode” for a specific odor.
Spi-Fly processes data obtained from sensors rather than the sensors that detect odors themselves. In the simulation, sensor readings are transmitted to a hidden layer representing Kenyon cells; mutually inhibitory neurons create a sparse pattern for each odor. This pattern is then matched with the output layer containing labeled odors. The system uses a simple learning rule that strengthens connections activated together with the correct response; it does not require the backpropagation method commonly used in modern neural networks.
In tests using odor data from common gas sensors, Spi-Fly reached its highest performance for each odor after only three exposures; the system using backpropagation required approximately 70 exposures to reach the same level. When new odors were added in groups, Spi-Fly retained the old odors with almost no loss of accuracy, while the performance of the other system approached the level of random guessing. The algorithm also showed less performance degradation on neuromorphic chips that mimic the brain with limited memory.
However, all the experiments were conducted in simulation and with single, isolated odors. The outcome for mixed odors in real-world environments remains uncertain. Max stated that the model could represent several hundred odors, depending on the conditions. The next stage is planned to involve transferring Spi-Fly to real odor-sensing hardware and neuromorphic chips developed by researchers at TU Eindhoven and Kiel University.
Why it matters
This approach could change how systems used in food inspection, environmental monitoring and security screening operate by reducing the problems of retraining electronic noses from scratch for new tasks and losing their previous knowledge. In particular, learning from a small number of samples and maintaining performance with limited memory could affect hardware and data requirements for applications working with different odors. Its not relying on backpropagation also makes the method significant for low-resource systems such as neuromorphic chips. However, the results are still limited to simulation and single odors; it is not known how mixtures in real-world environments will affect the algorithm’s performance in distinguishing odors and preventing forgetting. For this reason, the next stage will be to test the method on real odor-sensing hardware and neuromorphic chips.