Abstract
During the last years, there has been a lot of interest in achieving somekind of complex reasoning using deep neural networks. To do that, models likeMemory Networks (MemNNs) have combined external memory storages and attentionmechanisms. These architectures, however, lack of more complex reasoningmechanisms that could allow, for instance, relational reasoning. RelationNetworks (RNs), on the other hand, have shown outstanding results in relationalreasoning tasks. Unfortunately, their computational cost grows quadraticallywith the number of memories, something prohibitive for larger problems. Tosolve these issues, we introduce the Working Memory Network, a MemNNarchitecture with a novel working memory storage and reasoning module. Ourmodel retains the relational reasoning abilities of the RN while reducing itscomputational complexity from quadratic to linear. We tested our model on thetext QA dataset bAbI and the visual QA dataset NLVR. In the jointly trainedbAbI-10k, we set a new state-of-the-art, achieving a mean error of less than0.5%. Moreover, a simple ensemble of two of our models solves all 20 tasks inthe joint version of the benchmark.