🐁🧠 Results are for an 80x85x100 µm^3 mouse cortex IBEAM-mSEM volume. The work is a collaboration between @GoogleAI & the Hess lab at @HHMIJanelia, who pioneered this vEM technique. More soon!.

We find an OOM reduction in splits while maintaining mergers low enough that the traced axons should be useful for automated analysis or serve as an excellent starting point for further manual corrections.

We focus on the most challenging type of reconstruction (axon tracing) and evaluate PATHFINDER on a scale that ensures statistical significance and lack of bias (over 4 m total path length, all axons in the volume).

FFN v1.5 segments the volume EM images. Then SENSE defines the agglomeration space. Finally, SHAPE guides an efficient combinatorial search for optimal, morphologically plausible reconstructions.

PATHFINDER is an AI system designed to address this. It uses specialized models (FFN v1.5, SENSE, SHAPE) to process data across progressively larger spatial scales, currently up to the range of tens of microns.

Connectomics has made huge strides in recent years. But we can acquire data far faster than we can analyze it. With even insect brains requiring tens of years of manual corrections, proofreading remains a critical and expensive bottleneck, and a mouse brain seems out of reach.

Wouldn't it be great if we could not only image large connectomic volumes, but also completely reconstruct them? And if a whole mouse brain project didn't cost billions?. With the PATHFINDER preprint (, we preview a future where it doesn't have to.

RT @GoogleAI: Today, in collaboration w/ colleagues at the Institute of Science and Technology Austria (ISTA), we report the first-ever met….

RT @GoogleAI: In collaboration with HHMI Janelia & Harvard, we introduce ZAPBench, a whole-brain activity dataset and benchmark with single….

RT @NewsFromGoogle: Google Research, @Harvard and @HHMIJanelia have created the Zebrafish Activity Prediction Benchmark (ZAPBench), which c….

+ special shout-out to @alexbchen who recorded the activity dataset ZAPBench is based on!.

Wonderful collaboration between Google Research, HHMI Janelia and Harvard, with @janmatthis @a1mmer @Chinasaurli @mer_petkova @korfffly @EngertLab @MishaAhrens @stardazed0 and many more.

Explore interactive visualizations, datasets, and code through our website:. Find more info in the papers:

🕸️ Last but not least -- the connectome for this specific 🐟 specimen is currently being reconstructed and will be available at a later date!.

🧪 We test a number of SOTA time-series forecasting models to provide baselines. We also explore forecasting activity directly in voxel space in a companion paper.

📈This dataset forms the core of the Zebrafish Activity Prediction Benchmark (ZAPBench), which uniquely measures progress on forecasting neural activity at full brain scale and single cell resolution in a vertebrate.

🔬We collected and extensively processed a 4d dataset imaged with a lightsheet microscope. The resulting 3d movie covers over 70,000 neurons of a fish exposed to various visual stimuli.

🧠 How accurately can future neural activity be predicted from past activity in a whole brain? Larval zebrafish offer a unique opportunity to address this question. They are currently the only vertebrate species in which whole-🧠 activity can be recorded at cellular resolution.

Can we predict future brain activity in a small vertebrate?🤔. We're releasing ZAPBench⚡️(#ICLR2025 spotlight): a benchmark to forecast activity in a whole vertebrate brain🧠 at single-cell resolution!🐟 70k+ neurons, 3 TB of data & extensive baselines. Connectome is coming!.🧵👇

RT @alex_shapsoncoe: Our human connectome project is now published; Thanks to all contibutors, esp @michalj, D.Ber….