Skip-Frame Embeddings for Feature Adaptation and Visualization, Zain Shah

We present an unsupervised method for visualizing the generalization and adaptation capabilities of pre-trained features on video. Like the skip-grams method for unsupervised learning of word vector representations, we exploit temporal continuity in the target media, namely that neighboring frames are qualitatively similar. By enforcing this continuity in the adapted feature space we can adapt features to a new target task, like house price prediction, without supervision. The domain-specific embeddings can be easily visualized for qualitative introspection and evaluation.