Optimizing a Specialized Convolutional Neural Network in a Supercomputing Environment Using Tensorflow and VisIt.

We describe hyper-parameter optimization and training for a novel convolutional neural network that identifies neural axons in 3D volume electron microscope data for a larval zebrafish. Training a convolutional neural network to detect image components in three dimensional space has a variety of applications. There are many parts of a network that can be tuned to work most effectively in a supercomputing environment. We optimized run parameters including batch sizes, thread counts, and epoch sizes as well as hyper-parameters such as drop-out rates. Once we had improved the speed and accuracy of the network we created a large three dimensional volume to visualize the errors of the network. We added new samples to our training set based on the larger volume and improved the accuracy of the network using those samples. Adapting the network to our supercomputing environment and adjusting the dropout layers increased the accuracy and efficiency of the network. Being able to accurately detect neurons in three dimensional space shows potential for other deep learning applications.