The design of the proposed DHN is b

The design of the proposed DHN is based on two ideas.First, the network should deal with distance matrices whose size varies over time. Second, the receptive field of all elements of A t should be the whole D t , since the decision of the optimal assignment is global. While a fully convolutional approach could account for the first issue, large input matrices would imply partial receptive fields: the decision would be local and not global. The alternative is to use bi-directional recurrent neural networks (BiRNN).