Workshops Archive

Distributed Volumetric Generation of Neuronal Morphology at Scale


Workshop: AI4S: 8th Workshop on Artificial Intelligence and Machine Learning for Scientific Applications

Authors: Jyotsna Rajaraman (University of Illinois at Chicago) and Thomas Uram, Mark Hereld, and Michael E. Papka (Argonne National Laboratory)

Abstract: Large-scale electron microscopy connectomes enable the study of neuronal morphology, but annotating and proofreading these datasets remains effort-intensive. Generative models offer a complementary approach, learning from manual reconstructions to synthesize neuronal structures. Dense volumetric occupancy grids are directly compatible with connectomics workflows, but cost grows cubically with linear resolution which is unfavorable for sparse objects like neurons. We present a distributed generative pipeline with this representation and characterize its cost–fidelity tradeoff. We extract 46,582 axonal and dendritic fragments under a fixed 20 µm field of view and train a compartment-conditioned 3D convolutional variational autoencoder at different resolutions. Generated fragments are evaluated using geometric and topological descriptors calibrated against natural variation in the data. Across up to 256 NVIDIA A100 GPUs on Polaris, our pipeline retain 74–85% strong-scaling efficiency for training and generation is sharded to produce one million fragments in 3.1 s at 64³ and 5.25 min at 256³.


Back to AI4S: 8th Workshop on Artificial Intelligence and Machine Learning for Scientific Applications Archive Listing

Back to Full Workshop Archive Listing