Analysis of Memory Organization for Dynamic Neural Networks

Ying Ma, Jose Principe

Abstract: An increasing number of neural memory networks have been developed, leading to the need for a systematic approach to analyze and compare their underlying memory capabilities. Thus, in this paper, we propose a taxonomy for four popular dynamic models: vanilla recurrent neural network, long short-term memory, neural stack and neural RAM and their variants. Based on this taxonomy, we create a framework to analyze memory organization and then compare these network architectures. This analysis elucidates how different mapping functions capture the information in the past of the input, and helps to open the dynamic neural network black box from the perspective of memory usage. Four representative tasks that would fit optimally the characteristics of each memory network are carefully selected to show each network's expressive power. We also discuss how to use this taxonomy to help users select the most parsimonious type of memory network for a specific task. Two natural language processing applications are used to evaluate the methodology in a realistic setting.

Keywords: memory analysis, recurrent neural network, LSTM, neural Turing machine, neural stack, differentiable neural computers