Abstract
We present the winning solution for the Inclusive Images Competitionorganized as part of the Conference on Neural Information Processing Systems(NeurIPS 2018) Competition Track. The competition was organized to study waysto cope with domain shift in image processing, specifically geographical shift:the training and two test sets in the competition had different geographicaldistributions. Our solution has proven to be relatively straightforward andsimple: it is an ensemble of several CNNs where only the last layer isfine-tuned with the help of a small labeled set of tuning labels made availableby the organizers. We believe that while domain shift remains a formidableproblem, our approach opens up new possibilities for alleviating this problemin practice, where small labeled datasets from the target domain are usuallyeither available or can be obtained and labeled cheaply.