Researchers create Spi-Fly spiking network that learns odors from few examples
Researchers at the Okinawa Institute of Science and Technology, or OIST, introduced Spi-Fly, a spiking neural network inspired by the fruit fly olfactory system. The algorithm uses 1,000 hidden neurons and learns to classify odors after only a few exposures. On a gas‑sensor dataset of 200 recordings of ten compounds, Spi-Fly reached peak performance after each odor was seen just three times. A…
Key points
- Spi-Fly learns each odor after only three exposures on a 200‑sample gas‑sensor dataset.
- The network uses 1,000 hidden neurons and retains performance with 6‑bit weight precision, outperforming BPTT at 4 bits.
- Off‑chip memory needs are about five orders of magnitude lower than backpropagation methods.
Spi-Fly also demonstrated strong continual‑learning ability, adding new odor classes without significant forgetting, unlike a spiking network trained with backpropagation through time (BPTT). The model remained effective at 6‑bit weight precision and even outperformed BPTT at 4 bits. Researchers estimated its off‑chip memory requirements to be about five orders of magnitude lower than BPTT‑based methods, making it suitable for compact, energy‑constrained hardware. Limitations include lower accuracy after many training cycles, no updates to input‑hidden connections, and reduced performance with very low precision. The team plans to integrate Spi-Fly with odor‑sensing hardware alongside collaborators at TU Eindhoven and Kiel University.
Fruit fly-inspired AI learns smells quickly with far less memory
thebrighterside.news · 22 September 2026
Loading the full article…
This text was published by thebrighterside.news and written by Shy Cohen, Catherine Hodges. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Reddit discussionreddit.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Researchers identify Context poisoning as Extreme-Value attention interference in Long-Context language models · 1 src
- AdaMem improves soft compression for retrieval-augmented generation · 1 src
- Researchers propose CLIC to compare coding behaviors of large language models · 1 src
- Climate campaigns on X raise happiness by 9.02 points, reduce action language · 1 src
- Researchers introduce SJR architecture that improves multimodal moderation by 23.6% F1 · 1 src
Comments
via GitHub Discussions