Anchor & Transform: Learning Sparse Representations of Discrete Objects

Learning continuous representations of discrete objects such as text, sentences, users, and movies lies at the heart of many applications including involving text and user modeling. Unfortunately, traditional methods that embed all objects do not scale to large vocabulary sizes and embedding dimensions. In this paper, we propose a general method, Anchor & Transform (ANT) that learns sparse representations of discrete objects by jointly learning a small set of \textit{anchor embeddings} and a \textit{sparse transformation} from anchor objects to all objects. ANT is scalable, flexible, end-to-end trainable, and allows the user to easily incorporate domain knowledge about object relationships. ANT also recovers several task-specific baselines under certain structural assumptions on the anchor embeddings and transformation matrices. On several benchmarks involving text and user modeling, ANT demonstrates strong performance with respect to accuracy and sparsity.

• Machine Intelligence