We have developed deep learning algorithms to detect and classify pulsating auroras in video data captured by the Time History of Events and Macroscale Interactions during Substorms (THEMIS) All-Sky Imagers (ASIs). We label auroral data into four categories: “bad viewing condition”, “no aurora”, “other aurora” and “pulsating aurora”. In contrast to all prior studies centered on auroral image classification, our primary goal revolves around the classification of pulsating aurora. We introduce two distinct deep learning approaches: first, a convolutional neural network (CNN) model for single-frame classification combined with a smoothing algorithm; second, a hybrid model that combines a CNN with a recurrent neural network (RNN), meaning we are allowing temporal information to inform our classifications. Our models are trained on a large dataset comprised of 100,000 images classified by expert auroral observers. We tested our algorithms on a new dataset with 58 full-night videos and found real world accuracy values of 63.9% for the CNN method and 55.9% for the RNN method. An important outcome of this work is that we used our techniques to classify every one of the more than one billion images that comprise the THEMIS-ASI dataset and thereby produced the largest dataset of automatically classified pulsating aurora to date. Further, this work sets the stage for assimilation of auroral machine-learning outputs into geospace simulations.