Improving the Learning Platform for the Leukocoria Detection Project

Abstract

ABSTRACT Improving the Learning Platform for the Leukocoria Detection Project James Boer, M.S. Mentor: Gregory J. Hamerly, Ph.D. In this paper we describe a new and signi cantly improved learning platform for the leukocoria detection project. We have developed an easily maintainable and extensible system that can train hundreds of semi-random convolutional neural networks very quickly on GPUs. Docker is used to enable simple and platform-agnostic deployment of the system. The Torch machine learning library is used to train the convolutional neural networks and a PostgreSQL database is used to store the training results. A literature review of recent convolutional neural network research is done to nd methods of improving accuracy. Training results are promising, with single-network performance at 96.6% using minimal data and 99.5% using data augmentation