ExPrune is a general dynamic pruning optimization technique that enables partial computation of multi-grain rates for each input, without changing the model architecture or learning algorithm. It leverages the statistical property of exchangeability to identify relationships between some model parameters and intermediate values, and dynamically evaluates the network based on partial network evaluations to make pruning decisions. ExPrune has been applied to computer vision, graph models, and language models, demonstrating minimal computational effort and accuracy loss. It can also be used in conjunction with static-size pruning.
