Digital Fruit Flies Avoid 'Catastrophic Forgetting' by Refusing to Learn

The New Intelligence · Explore this series ›

Tormod Guldvog

Editor & Duowriter

September 10, 2026

Key Takeaways

• A fly-inspired spiking network resists catastrophic forgetting by freezing almost every weight.
• Only output weights train, so no class can overwrite another.
• It learns in three presentations where backpropagation needs about seventy.

✓ Fact-checked · See verification details

AI assistant: Research, Drafting, QA

A fruit fly smells something, and about a hundred of its neurons go quiet.

That is roughly the shape of the trick. In the mushroom body, the paired structure where a fly stores what it has learned about odors, sit around 2,000 Kenyon cells per hemisphere. Each one samples input from a small, apparently random handful of the fifty or so projection neuron types carrying signal up from the antenna.

A Columbia group including Richard Axel, who shared a Nobel Prize in 2004 for working out how the sense of smell is wired, traced 200 of those cells in 2013 and found no organizing principle at all: the wiring showed "no discernible organization with respect to their odour tuning, anatomic features or developmental origins."

Then one enormous inhibitory neuron, the anterior paired lateral cell, one per hemisphere, shouts down about 95 percent of the Kenyon cells. What survives is a sparse pattern, five percent of the population, different for every smell. A bar code, in a way.

In 2014 Andrew Lin, Gero Miesenböck and their colleagues at Oxford's Centre for Neural Circuits and Behaviour cut the feedback loop between the Kenyon cells and the APL neuron and watched what happened to the animal. Disrupting that loop, they reported, "decreased the sparseness of Kenyon cell odor responses, increased inter-odor correlations and prevented flies from learning to discriminate similar, but not dissimilar, odors."

Break the silence and the fly can still tell vinegar from smoke; it can no longer tell one alcohol from its near neighbor.

A Silicon Fly

Kevin Max and Yang Shen, both at the Okinawa Institute of Science and Technology, have built a spiking neural network out of that arrangement and called it Spi-Fly. Shen brings an unusual pedigree to it: she was first author, at Cold Spring Harbor Laboratory with Saket Navlakha and with Sanjoy Dasgupta at UC San Diego, of the 2023 model that turned the fly's mushroom body into a defense against catastrophic forgetting, and which Spi-Fly builds on directly. Spi-Fly is that model carried across the world and rebuilt for silicon.

The architecture is quite simple. Seventy-two sensor inputs (or five, on the second dataset) fan out to a hidden layer of a thousand neurons through connections that are random, sparse, and permanently frozen. Mutual inhibition among those thousand leaves about five percent firing. Only the wires from that surviving five percent to a bank of output neurons are ever allowed to change. Each odor class gets an output neuron entirely to itself.

The results were reported in Neuromorphic Computing and Engineering in August.

On a gas sensor dataset of ten industrial compounds, Spi-Fly reached its best accuracy after three presentations of each odor. Backpropagation through time, the standard method for training a network that unfolds in time, needed about seventy to get there.

Train Spi-Fly on two odor classes, then two more, then two more, testing each time on everything it has seen, and it holds. Backpropagation, run through the same gauntlet, collapses to barely above chance. Elastic weight consolidation, the best-known fix for exactly this problem, did not rescue it.

That is catastrophic forgetting, and Spi-Fly is immune to it.

Nothing in the network can interfere with anything else. Learning class 4 adjusts only the weights leading to output neuron 4. Output neuron 1 is untouched, by construction, because there is no shared plastic parameter for the new lesson to overwrite. Max and Shen say so without embarrassment: "learning of one class therefore does not change weights responsible for identification of another class."

"This is expected," they write, "since the Spi-Fly model does not perform representation learning as BPTT (there is no learning of the input-to-hidden weights)."

The first layer never trains. It is a random projection that was fixed before the first odor arrived and stays fixed after the last one, so there is nothing in it for experience to improve. On the synthetic Drosophila dataset, Spi-Fly plateaus at 58.9 percent and stays there no matter how much more data you feed it. Backpropagation, slow and forgetful, climbs past it to 83.9.

And a Gaussian Naive Bayes classifier, a method a statistics student meets in about week three, scores 93.9 on the same task.

The authors give a fair reason for it: the offline classifiers see the whole dataset at once and are not obliged to treat time as an input dimension.

This does not mean that copying a fly beats conventional machine learning at telling smells apart. What has been shown is that a network which almost never changes learns fast and does not forget, which is a different and more interesting claim, which belongs to a longer argument about whether brains compute the way software does.

Immunity to forgetting and inability to improve are one property, examined from either end.

Two cautions

The first is that nothing here has smelled anything. Spi-Fly ran entirely in simulation, in PyTorch, on prerecorded sensor archives and on synthetic traces generated from the authors' own earlier model. No neuromorphic chip was involved and no live sensor sampled any air. The paper's looks forward: "We plan to implement Spi-Fly on our neuromorphic olfactory sensing system to test its performance under real-life conditions."

The test sets are also small: 40 samples on the gas sensor task, 90 on the synthetic one. On 40 samples a single misclassification moves the score by 2.5 points, which is a useful thing to hold in mind next to a figure quoted to one decimal place.

The second caution comes from Thomas Cleland, professor of psychology at Cornell, who co-designed the EPL net that Spi-Fly benchmarks itself against and beats. He told Ars Technica that the two models "are based on different layers of the olfactory system and are built for different noise models," his EPL net drawn from the mammalian olfactory bulb rather than the insect mushroom body. He is not a disinterested party. He is also the person best placed to say it, and the comparison is the weakest link in the paper.

Max, for his part, is candid about the ceiling. Asked how many odors his model could hold, he told the same reporter it should manage "a couple hundred," depending on how similar they are to each other. That is a long way from never forgetting anything.

More on brain research

A Brain Cell That Always Knows Which Way You're Facing

Scientists found a neuron that tracks your direction whether you're sitting still or running–and losing it may explain Alzheimer's disorientation.

→

We tend to treat catastrophic forgetting as a bug awaiting a smarter optimizer. Spi-Fly suggests it may be a price, and that the currency it is paid in is plasticity. You can build a system that holds everything it has been told, provided you build one that cannot get better at holding it. The fly, working with 2,000 cells and no gradient descent, seems to have taken that bargain a very long time ago and to be doing perfectly well on the terms.

Whether the bargain is worth taking in silicon is an engineering question, and it is unanswered, because nobody has yet run this on a chip. But the shape of it is now clear enough to argue about, which is more than most brain-inspired architectures manage.

Sources

• Primary source: Max, Kevin & Yang Shen. "Few-shot, continual learning for spiking neuromorphic olfaction." Neuromorphic Computing and Engineering 6, 034013 (2026). Open access, CC-BY 4.0.
• Context sources:
  • Caron, Sophie J. C., Vanessa Ruta, L. F. Abbott & Richard Axel. "Random convergence of olfactory inputs in the Drosophila mushroom body." Nature 497, 113-117 (2013).
  • Lin, Andrew C., Alexei M. Bygrave, Alix de Calignon, Tzumin Lee & Gero Miesenböck. "Sparse, decorrelated odor coding in the mushroom body enhances learned odor discrimination." Nature Neuroscience 17, 559-568 (2014).
  • Shen, Yang, Sanjoy Dasgupta & Saket Navlakha. "Reducing Catastrophic Forgetting With Associative Learning: A Lesson From Fruit Flies." Neural Computation 35, 1797-1819 (2023).
  • Imam, Nabil & Thomas A. Cleland. "Rapid online learning and robust recall in a neuromorphic olfactory circuit." Nature Machine Intelligence 2, 181-191 (2020).
  • Sgorbissa, Federica. "Just like a fruit fly, a new algorithm never forgets old scents." Ars Technica, 3 September 2026. Source of the Cleland and Max email quotes.
  • Hodges, Catherine. "How to decipher smells like a fruit fly." OIST News Center, 24 August 2026.

Tormod Guldvog

Editor & Duowriter

Tormod Guldvog is the editor of Science Reader, covering the space where human and artificial intelligence meet. With a background in technology and science communication, he writes about how AI is changing scientific discovery — and what that means for us. More about Science Reader

Cite this article

Tormod Guldvog. (2026, September 10). Digital Fruit Flies Avoid 'Catastrophic Forgetting' by Refusing to Learn. Science Reader. https://sciencereader.com/catastrophic-forgetting-fruit-fly-network/