AI-Cafe presents: Deep Ensembles for the Prediction of Media Interestingness

In the context of the ever-growing quantity of multimedia content from social, news and educational platforms, generating meaningful recommendations, ratings, and filters now requires a more advanced understanding of their impact on the user, such as their subjective perception. Visual interestingness is one of the most important subjective perception concepts and is currently a popular avenue of research in affective computing.

However, given the high degree of complexity of this subject caused by its inherent multimodality and subjectivity, classical single-system deep neural network-based approaches currently show relatively low prediction capabilities when compared with other computer vision concepts. We will therefore present our advances in deploying a set of larger ensembling-based architectures that use late fusion for greatly increasing single-system results. This session will present one of the first attempts at using a deep neural network as the primary ensembling engine, as well as introduce some novel neural network layers and architectures that are specially designed for ensembling.