EDNC: Evolving Differentiable Neural Computers

Abstract

One recent study in the field of deep learning is extending Artificial Neural Networks (ANNs) by coupling them to external memory resources. Neural Turing Machine (NTM) and Differentiable Neural Computer (DNC) are two counterparts in this field. Research activities fall into two categories of either interacting with memory or choosing controller components. The first approach uses a try and error fashion to provide the controller’s structure that would be a problem when there is no prior knowledge on the controller’s structure for particular tasks. The second approach includes methods for choosing the controller’s structure and weight optimization. NeuroEvolution falls in the second category that automatically and without prior knowledge obtains an appropriate network structure. This research presents Evolutionary Differentiable Neural Computer (EDNC), which uses a novel NeuroEvolution algorithm that is introduced in two types of nested object-oriented encoding called Adaptive Layer NeuroEvolution (ALNE) and Matrix-based one called M_ALNE. These NeuroEvolution algorithms use the following sub-algorithms: Self-Adaptivity in the Initial Population (SAIP), Self-Adaptivity in Evolutionary Structures (SAES), Layer Recombination (LR), Layer Mutation (LM), and Self-Adaptivity in Mutation (SAM). The evolution process starts with SAIP, and then it takes short and long steps to explore a variety of neural structures by SAES algorithm that uses both LR and LM algorithms to provide a variety of structures. Finally, the SAM algorithm guides the selected structures to a target structure. In this research, experiments applied EDNC on well-known tasks, including Facebook bAbI, copy, graph (shortest path), and 8-puzzle. The experiments show that the proposed method in both encodings automatically and without prior knowledge, produces an appropriate controller structure in the shortest time. It though shows that both proposed encodings reduce the evolution time by at least 73% in comparison with the baseline methods.