Interpretable Structure Induction Via Sparse Attention

摘要

Building upon the principle of parsimony, we propose sparse, regularized, constrained and structured hidden layers. We seek to discuss the potentials of these strategies with an expert community on black-box interpretability.