Deep neural networks can produce embeddings, structured representations of unstructured data, which can be queried, indexed, and compared like structured data. In this post, Peter Gao (CEO of Aquarium Learn) provides an overview of neural network embeddings, which are learned during supervised training as an intermediate representation. He discusses how to reduce and visualize embeddings to explore interesting clusters or outliers and examines how embeddings can be applied to facilitate model debugging and identify labeling errors.