Sharing open deep learning models


Dalgali A., Crowston K.

52nd Annual Hawaii International Conference on System Sciences, HICSS 2019, Hawaii, United States Of America, 8 - 11 January 2019, vol.2019-January, pp.2109-2118 identifier

  • Publication Type: Conference Paper / Full Text
  • Volume: 2019-January
  • City: Hawaii
  • Country: United States Of America
  • Page Numbers: pp.2109-2118
  • Yıldız Technical University Affiliated: No

Abstract

We examine how and why trained deep learning (DL) models are shared, and by whom, and why some developers share their models while others do not. Prior research has examined sharing of data and software code, but DL models are a hybrid of the two. The results from a Qualtrics survey administered to GitHub users and academics who publish on DL show that a diverse population shares DL models, from students to computer/data scientists. We find that motivations for sharing include: increasing citation rates; contributing to the collaboration of developing new DL models; encouraging to reuse; establishing a good reputation; receiving feedback to improve the model; and personal enjoyment. Reasons for not sharing include: lack of time; thinking that their models would not be interesting for others; and not having permission for sharing. The study contributes to our understanding of motivations for participating in a novel form of peer-production.