The recent success of Deep Neural Networks (DNNs) for object recognition has opened a new avenue for computational models of visual attention due to the tight link between visual attention and object recognition. In this talk, I will show that DNNs for object recognition can be adapted to predict properties of the eye-fixations such as their location, temporal sequence and consistency among subjects. Results demonstrate the emergence of saliency-selective features in the DNN which are distinct from object-selective features. Also, results suggest that the eye-fixation consistency among subjects is an attribute of natural images that can be predicted from the DNN features. Our model significantly outperforms previous approaches and leads to state-of-the-art models.