Structured Memory for Neural Turing Machines

• DOI: 10.48550/arxiv.1510.03931
• Publication Date: 2015-01-01
• Authors: Wei Zhang; Yu Yang; Bowen Zhou
• Keywords: Component (thermodynamics); Memory model; Turing; Memory map

Neural Turing Machines (NTM) contain memory component that simulates "working memory" in the brain to store and retrieve information ease simple algorithms learning. So far, only linearly organized is proposed, during experiments, we observed model does not always converge, overfits easily when handling certain tasks. We think key some faulty behaviors of NTM, better organization could help fight those problems. In this paper, propose several different structures for proved experiments two our proposed structured-memory NTMs lead convergence, term speed prediction accuracy on copy task associative recall as (Graves et al. 2014).