Google’s AI Is Generating Fake Neurons to Speed Up Real Brain Mapping

Brain mapping is one of those problems where brute force just doesn’t scale. The recently released fruit fly connectome, with 166,000 neurons, took years of combined AI and human effort. A mouse brain is a thousand times bigger than that. A human brain is a thousand times bigger still. You can’t just throw more compute at it and expect clean results.

Google Research has been chipping away at this for over a decade through their Connectomics team. Their latest trick, presented at ICLR 2026, is a model called MoGen (Neuronal Morphology Generation) that creates synthetic neuron geometries. The idea is simple: train your AI on more varied neuron shapes, and it’ll get better at recognizing real ones. The results are a 4.4% reduction in reconstruction errors. That sounds small, but at the scale of a complete mouse brain, it translates to roughly 157 person-years of manual proofreading saved. I’ll take that trade.

The core problem in connectomics is that neurons look nothing like typical cells. Most cells are roughly spherical blobs. Neurons are spindly, branching messes with long axons that curl and twist, dendrites covered in tiny spines, and thousands of synaptic junctions. Their shape directly relates to their function, which is why getting the reconstruction right matters. Google’s previous model, PATHFINDER, identifies neurite segments and stitches them together, but it struggles with the sheer diversity of neural morphologies.

MoGen addresses this by generating synthetic neurons through point cloud flow matching. You start with a random cloud of points and iteratively morph it into a realistic neural shape. The process is surprisingly elegant to watch — the animation in the blog post shows those initial noise blobs gradually resolving into recognizable neurons with axons and dendrites. Training on these synthetic shapes alongside real data makes the downstream reconstruction model more robust.

What I find interesting is that this isn’t about building a better segmentation model from scratch. It’s about augmenting the training data in a smarter way. The Connectomics team has been doing this long enough to know that the bottleneck isn’t the initial AI pass — it’s the manual proofreading that catches the remaining errors. Shave off a few percentage points of error, and you save months of human labor per brain.

Of course, we’re still years away from a full mouse brain map, let alone a human one. But this approach feels more practical than waiting for some breakthrough in imaging or compute. Synthetic data augmentation has worked in other domains, from autonomous driving to medical imaging. Seeing it applied to connectomics makes sense.

The paper is worth a read if you’re into the technical details. MoGen is available as a model, and the team has put it up alongside their other connectomics tools. I suspect we’ll see more groups adopting similar strategies as brain mapping scales up.