In this talk we touch on several problems in machine learning that can benefit from the use of topic modeling that stem from large scale data. We present topic modeling based approaches for online prediction problems, computer vision, text generation, and others. While these problems have classical machine learning approaches that work well, we show that by incorporating contextual information via topic features, we obtain enhanced and more realistic results. This problem serves as a motivating application that demonstrates how new mathematical methods are needed to understand large-scale data. In this talk we provide a brief overview of these problems and show how topic features can be used in these settings. We include supporting theoretical and experimental evidence that showcases the broad use of our approaches.