Abstract
Determinantal point processes (D PPs) provide an elegant and versatile way to sample sets of items that balance the quality with the diversity of selected items. For this reason, they have gained prominence in many machine learning applications that rely on subset selection. However, sampling from a D PP over a ground set of size N is a costly operation, requiring in general an preprocessing cost and an sampling cost for subsets of size k. We approach this problem by introducing D PP N ETs: generative deep models that produce D PP-like samples for arbitrary ground sets. We develop an inhibitive attention mechanism based on transformer networks that captures a notion of dissimilarity between feature vectors. We show theoretically that such an approximation is sensible as it maintains the guarantees of inhibition or dissimilarity that makes D PPs so powerful and unique. Empirically, we show across multiple datasets that D PP N ET is orders of magnitude faster than competing approaches for D PP sampling, while generating high-likelihood samples and performing as well as D PPs on downstream tasks.