Fast-weight continual learning in a fly

LLMs have already started to revolutionize math. I wanted to see if they can also make discoveries in systems neuroscience, a field I left 2+ years ago. Over two days, Nico and I pointed Fable 5.1 at an open problem: finding the neural circuits that perform path integration. We intentionally stayed ~2 levels above an experimentalist and let the model read the literature, screen the fly’s complete wiring diagram without assuming which cells matter, and simulate the best candidates.

We arrived at 4 findings. The main one is that four neuron types, hΔH, hΔA, hΔI and hΔG, are capable of storing the running sum of the fly’s steps in the strength of their synapses. If true, this would be fast-weight continual learning, the same mechanism as linear-attention and state-space models, whose state is a running sum of gated writes into a weight matrix. After sharing these results on twitter, we found that earlier work¹ had independently proposed the same neurons, and two recent dissertations^2,3 show that two of them, hΔG and hΔA, indeed perform path integration, though it is unclear whether integration occurs through activity or through weights.

The question

The navigational system of the fruit fly is a crown jewel of systems neuroscience due to the work of some seminal neuroscientists (Larry Abbott, Gaby Maimon, Vivek Jayaraman, Barbara Webb), but it is an unfinished story.

A fly that leaves a drop of food and wanders in the dark can always find its way back. To do that it has to keep a running sum of every step it has taken, a process called path integration.

Path integration needs four things: the speed of each step, the direction of each step, a place to add the steps up, and a way to turn the sum into a direction home. To add up the steps, a brain can hold the running sum in two ways:

1. In activity. Neurons keep firing, exciting each other in a loop tuned so precisely that the signal neither fades nor blows up. Most models of navigation assume this, and it is how RNNs hold state.
2. In weights. Nothing keeps firing. Each step is written into synaptic strengths, and their sum is the sum of the journey.

We don’t know which one the fly uses, or which neurons do it.

Findings

We found candidate neurons at each stage of path integration.

1. Finding 1The speed of each stepWe traced the fly’s turning and walking speeds back through the connectome to two cell types nobody has recorded, PS196_b and FB3A.Read finding →
2. Finding 2The direction of each stepWe show that the neurons that turn the fly’s compass get most of their input from the cells they drive, a loop the standard models leave out. In our model, this loop stops the compass from turning.Read finding →
3. Finding 3 · main findingA place to add the steps upFour cell types are wired to store the sum of the fly’s steps in their synapses. The best, hΔH, and a weaker one, hΔI, have never been recorded. In simulation, hΔH’s wiring can store most of the path.Read finding →
4. Finding 4A way to turn the sum into a direction homeAs a fly walks away from food, it stores a vector pointing from the food to itself. To walk back, something must flip that memory. We propose two cell types, hΔM and hΔI, wired to do this before the memory reaches the neurons that steer.Read finding →

Convergent with experiments

After we posted, others pointed us to earlier work. Prior papers, including Maimon and Abbott¹, had already proposed the hΔ neurons as path integrators, and two dissertations^2,3 had already run experiments on two of them, showing that hΔG and hΔA add up the fly’s steps over a few seconds.

So two independent routes reached the same neurons: earlier work through hypothesis-driven study, and us through an unbiased screen of the whole connectome, run largely by an LLM in two days. Both dissertations are still embargoed, so whether path integration runs on synaptic updates, persistent activity or both is still open. We tried to cover the literature exhaustively and certainly missed some of it, so corrections are welcome.

Using AI to do science

The good: Fable 5.1 was creative, took a lot of scientific paths that were in high taste, and most of the time found good sanity checks. It got me up to speed on a lot of the research and performed huge, unbiased screens of connectomics data. It is really not emphasized enough how far along Fable’s research and scientific capabilities are, probably because it can’t be felt by making three.js games.

The bad: it needed a lot of pushing to go outside of what is currently known, and even then, it still clung on to existing hypotheses. There were a whole host of small factual errors, incomplete coverage of current research, misattribution, and terrible writing: it does not seem to have a theory of mind over what people know and don’t know. Lastly, Claude Code should compact at ~400k, not 1M. Responses go haywire to a point of no return after this.

The lessons appear to be the same as in software engineering: to do serious, careful work, verification is critical. Automated verification is hard in a lot of science, so it requires human expertise. I could not independently verify the completeness of published and unpublished results, nor check against long-context hallucinations. A serious systems neuroscientist would be able to flag these issues and extract far greater use from LLMs than I would, just like how a senior SWE with good systems understanding can be unlocked to a far greater extent than a meat proxy.