Finding ReMO (Related Memory Object): A Simple Neural Architecture for Text based Reasoning
Finding ReMO (Related Memory Object): A Simple Neural Architecture for Text based Reasoning
复制标题
寻找 ReMO（相关记忆对象）：基于文本推理的简单神经架构
DOI：
--
复制
发表时间：
2018
期刊：
影响因子：
--
通讯作者：
Sungzoon Cho
中科院分区：
文献类型：
--
作者：
Jihyung Moon;Hyochang Yang;Sungzoon Cho
基于记忆网络的模型在关系推理任务上取得了显著进展。最近，一种更简单但功能强大的神经网络模块被称为关系网络（RN）。尽管关系网络结构简单，但其时间复杂度随数据呈二次增长，限制了其在大规模内存任务中的应用。我们介绍相关记忆网络，一个端到端的神经网络架构，利用记忆网络和关系网络结构。我们遵循记忆网络的四个组成部分，每个组成部分的操作与关系网络相似，而不需要一对对象。因此，我们的模型与RN一样简单，但计算复杂度降低到线性时间。它在bAbI-10k基于故事的问答和bAbI对话数据集的联合训练中取得了最先进的结果。
Memory Network based models have shown a remarkable progress on the task of relational reasoning. Recently, a simpler yet powerful neural network module called Relation Network (RN) has been introduced. Despite its architectural simplicity, the time complexity of relation network grows quadratically with data, hence limiting its application to tasks with a large-scaled memory. We introduce Related Memory Network, an end-to-end neural network architecture exploiting both memory network and relation network structures. We follow memory network's four components while each component operates similar to the relation network without taking a pair of objects. As a result, our model is as simple as RN but the computational complexity is reduced to linear time. It achieves the state-of-the-art results in jointly trained bAbI-10k story-based question answering and bAbI dialog dataset.