In order to efficiently study the brain through microscopy, it is necessary to have robust methods for discovering and labeling structures of interest. Deep learning tools have been successfully applied to a number of neuroscientific studies, however in many cases trained neural network models are insufficient on their own. In particular, learning-free methods can enable semi-automated annotation, greatly increasing the amount of available training data, and hyperparameter optimization can discover neural network models better suited for processing neural volumes. This presentation will introduce two tools for supplementing and improving deep learning for neuroscience: the Flexible Learning-Free Reconstruction of Neural Volumes pipeline (FLoRIN), and the Scalable Hardware-Aware Distributed Hyperparameter Optimizer (SHADHO). These open-source tools have been deployed at scale on Argonne Leadership Computing Facility resources to process volumes imaged by the Advanced Photon Source.

Bio

Jeff Kinnison is a graduate research assistant in the Computer Vision Research Lab at the University of Notre Dame. He has been a guest graduate researcher with Argonne National Laboratory with the Biosciences division since August 2016. His research interests include hyperparameter optimization for machine learning, distributed computing, and applications of (learning and learning-free) computer vision to neuroscience.