AI-generated synthetic neurons are making brain mapping less painful

The Google Research Connectomics team has been at this for over a decade, and their latest trick is using AI to generate fake neurons that help real brain mapping go faster.

Their new paper, “MoGen: Detailed neuronal morphology generation via point cloud flow matching,” is set to appear at ICLR 2026. The idea is straightforward: train a model to produce synthetic neuron shapes, then use those as extra training data for the AI that actually reconstructs neurons from microscope images.

And it works. They report a 4.4% reduction in reconstruction errors. That sounds like a small number, but at the scale of a complete mouse brain — which is a thousand times larger than the fruit fly brain they just mapped — it translates to 157 person-years of manual proofreading saved. That’s not nothing.

The fruit fly brain map they released recently had 166,000 neurons and took years of AI-assisted work plus human experts. A mouse brain would be 166 million neurons. A human brain? 166 billion. We’re not getting there with brute force.

Connectomics is the field that tries to create wiring diagrams of brains. It starts with slicing brain tissue into thin sheets, imaging them, then stacking and aligning the 2D images to reconstruct 3D neurons. The first complete worm brain map took 16 years of manual labor. Modern methods use digital imaging and AI, but the final proofreading step is still done by humans. That’s the bottleneck.

Neurons are weird cells. Most cells in your body are roughly spherical. Neurons are spindly, branching things with axons that can be long and twisty, dendrites that look like tangled tree roots, and tiny spines that form synapses. Their shape matters for function, and that’s what makes reconstruction hard.

Google’s previous model, PATHFINDER, works by identifying neurite segments — subsections of a neuron — and then stitching them together. The problem is that real neurons have so much variety that the training data never covers all the edge cases. Synthetic neurons fill in those gaps.

MoGen uses point cloud flow matching to generate realistic morphologies. The animation they released shows point clouds gradually resolving into recognizable neural shapes. It’s trained on mouse neurons, and the results look convincing enough to fool the reconstruction model into learning better.

This isn’t the first time synthetic data has been used to augment training sets, but it’s one of the more clever applications I’ve seen. The key insight is that you can generate infinite variations of neuron shapes, which is exactly what a deep learning model needs to generalize well.

Is 4.4% going to change the world overnight? No. But compound that across the decade-long effort to map larger brains, and it starts to add up. The Connectomics team has already mapped fragments of zebra finch brain, whole larval zebrafish brain, a small piece of human brain, and they just launched a mouse brain mapping project.

What I’d really like to see is how well MoGen generalizes across species. If it can produce plausible synthetic neurons for human brains without retraining from scratch, that would be a big deal. The paper only covers mouse neurons so far.

Also worth noting: this is a research project, not a product. Don’t expect to download MoGen and start generating your own synthetic neurons tomorrow. But the technique is sound, and it’s another step toward making brain mapping practical at scale.

One thing that bothers me about the press coverage is the focus on the error reduction percentage without context. A 4.4% improvement in reconstruction error is meaningful when your baseline is already good. If your reconstruction model is 95% accurate, 4.4% improvement brings you to 99.4% — that’s a huge difference in practical terms.

And that 157 person-years number? That’s real labor. Real people sitting at screens correcting AI mistakes. If MoGen can eliminate even a fraction of that, it’s worth paying attention to.

The broader trend here is that AI is becoming a tool for generating its own training data, which is a meta-level advance that I think is underappreciated. We’re moving from “train AI on human-labeled data” to “train AI on AI-generated data that’s good enough to train another AI.” That’s a paradigm shift.

Google Research has been quietly building a stack of foundational tools for connectomics over more than a decade. MoGen is the latest addition. It’s not flashy, but it’s solid engineering.

If you’re interested in the details, the paper is at ICLR 2026, and the MoGen model description is on their connectomics website. Worth a look if you care about how we’re going to map the brain without waiting another century.