Reviews: DppNet: Approximating Determinantal Point Processes with Deep Networks

This work focuses on sampling determinantal point processes (DPPs) by utilizing deep learning architecture so that the running time for sampling becomes faster than the standard DPP sampling. To avoid complex operations (e.g., matrix inversion or eigendecomposition) used in the standard DPP sampling, authors come up with an attention mechanism where it simply needs to dot-product of features or kernel matrix itself. As far as I know, most of the prior works in deep learning community borrow DPPs and apply them at the top of layer in the network for the purpose of generalization, compression and so on. Unlike the priors, this work utilizes deep learning models in order to simplify the DPP sampling. This can be a great impact for future works of other models whose inputs are a subset of items and guide them to capture their characteristic (e.g., ranking) using deep learning architectures.