Recent advances in unsupervised salient regions detection algorithms made possible to obtain high-quality saliency predictions without human annotated data. In this paper, we explore the possibilities of semi-supervised salient region predictions using neural networks. We built a fully-convolutional deep architecture and performed controlled experiments training the same architecture from the ground up while using differently generated data as labels. We show that efficient combination of multiple unsupervised saliency prediction algorithms has a consistently positive impact on the predictions generated by a deep model. Despite the increase in model performance, we show that supervised models are still vastly superior in terms of quality. © 2018 The Author(s).