UC San Diego

Inspired by the success of convolutional neural networks in image classification, and other higher level vision tasks, we explore two applications of such deep convolutional neural networks to model tasks typically involving human assessment, viz. i) prediction of personality from social media images, and ii) prediction of citations from the visual elements of an academic paper. The aim in this context is to discover if there is any predictable and learnable signal in the input data. As an extension, we attempt to discover what aspects of the signal are indeed learnt that lead to the results presented. For instance, if personality can be predicted, what aspects of the image are causing that? Similarly if an academic paper is highly cited, what are the characteristic visual elements that cause this? We employ convolutional neural networks in order to understand what imputable attributes we may derive that are simpler to reason.