Emotional state influences nearly every aspect of human cog-nition. However, coding emotional state is a costly processthat relies on proprietary software or the subjective judgmentsof trained raters, highlighting the need for a reliable, automaticmethod of recognizing and labeling emotional expression. Wedemonstrate that machine learning methods can approach near-human levels for categorization of facial expression in natural-istic experiments. Our results show relative success of modelson highly controlled stimuli and relative failure on less con-trolled images, emphasizing the need for real-world data forapplication to real-world experiments. We then test the poten-tial of combining multiple freely available datasets to broadlycategorize faces that vary across age, race, gender and photo-graphic quality.