Explainable Sparse Attention for Memory-Based Trajectory Predictors

Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, Alberto Del Bimbo

Abstract: In this paper we address the problem of trajectory prediction, focusing on memory-based models. Such methods are trained to collect a set of useful samples that can be retrieved and used at test time to condition predictions. We propose Explainable Sparse Attention (ESA), a module that can be seamlessly plugged-in into several existing memory-based state of the art predictors. ESA generates a sparse attention in memory, thus selecting a small subset of memory entries that are relevant for the observed trajectory. This enables an explanation of the model’s predictions with reference to previously observed training samples. Furthermore, we demonstrate significant improvements on three trajectory prediction datasets.