AI Learning Efficiency Boost in Simulation

A new algorithm mimics fruit fly olfaction to enable AI systems to retain historical knowledge during continuous learning, with significant efficiency gains in simulated odor classification.

Okinawa Institute of Science and Technology researchers published the Spi-Fly algorithm in Neuromorphic Computing and Engineering on September 3, 2026. The system draws inspiration from fruit fly olfactory pathways—using 2,000 Kenyon cells to implement sparse coding—achieving odor classification in just three exposures versus seventy required by traditional backpropagation methods. Theoretical capacity reaches 75 million odor barcodes (100 neurons selecting 5), though real-world performance is constrained to hundreds of odors due to noise and environmental factors. All testing occurred in simulation using isolated odors, with no validation in complex, multi-odor environments.

This efficiency gain addresses a key bottleneck in AI’s ability to retain historical knowledge during continuous learning. By reducing exposure cycles, Spi-Fly could lower computational demands for systems that require long-term memory—such as health monitoring or security applications where historical data patterns matter. However, the algorithm’s practical utility remains unproven in real-world settings. Current limitations include dependency on odor similarity and absence of environmental interference testing. Next steps involve validating performance in noisy, multi-odor scenarios to determine if the efficiency gains translate to tangible improvements in resource-intensive applications.

Source: Ars Technica