Computer Vision News - October 2020

2 Open Source Deep Learning Platform 36 Why be an Open Source Deep Learning Developer? by Stephen Aylward, Kitware, Inc. The field of deep learning has two dominant characteristics: outstanding performance and openness . While the outstanding performance of deep learning systems is well known and is rapidly spreading to new application areas, the openness of deep learning is perhaps less well publicized yet nevertheless equally important. The openness of deep learning refers to the dominance of open science practices in the field. It is standard practice to offer open publications, e.g., post on arXiv , to share data, e.g., ImageNet , and to use and contribute to open source deep learning platforms . Furthermore, for deep learning research and development, this openness spans academia and industry, where we see the same platforms being used in research and commercial development. These open platforms include well established, general purpose systems such as PyTorch and the Insight Toolkit (ITK) , as well as dynamic, cutting- edge, domain specific systems such as the Medical Open Network for AI (MONAI) platform . These systems allow the outstanding performance of deep learning systems to be readily shared, compared, commercialized, and extended. This dependence on open source software also represents a major change in attitude. The adage “you get what you pay for” no longer seems to be applied in this context. However, two major questions arise: Who is developing and maintaining the open source software of deep learning, and why do they do it? To answer that question, we interviewed several lead developers of and contributors to MONAI and ITK: Nic Ma (NVidia) , Eric Kerfoot (Kind’s Collect London) , Wenqi Li (NVidia) , Annika Reinke (German Cancer