Design Semantic Memory by Artificial Neural Network

Abstract

This paper designs a multi-layer neural network to simulate semantic memory [1]–[2] for building a hierarchical network model. The hierarchical network model [3] consists of nodes (concepts) and links, which connect between the nodes. Nevertheless, the distance (relevance) of link is defined by human and cannot be automatically adapted after the changing of environment. This study proposes an artificial neural network to learn the scenario examples, which generates the relevance (distance) between nodes in the hierarchical network. It can be also automatically adapted when the environment changes, which is similar to human memory re-encoding the semantic memory. There are three issues about the learning effect discussed: (1) different number of layers (three to six) suitable for the designed neural networks, (2) different activation functions (sigmoid, tanh(x), ReLU(x)) suitable for the neural unit, (3) the normalization and softmax [4] function suitable for the neural output and their effects on learning, and whether the distance generated by neural network automatically adjusts when the environment changes. The experimental results show that the proposed artificial neural network can successfully build the semantic memory hierarchical network model. When the environment changes, it can also re-learn automatically to adapt the network distance and generate a new semantic memory network model. We propose an innovative approach which could be widely applied to the fields of brain science, psychology and education.