A Study on Realizing Awareness Using 3VL-MLP

Abstract

Awareness is a way from sensory data to cognition. The main purpose of computational awareness (CA) is to understand the awareness mechanism and realize it in computers. Various awareness are used in our daily lives for making decisions, but most of them are tacit. For the purpose of CA, we need to interpret and understand tacit awareness as far as possible. In our earlier study, we introduced a general graph model of aware systems. In this paper, we focus on the multilayer perceptron (MLP) model, and study the feasibility of interpreting MLPs using 3-valued logic (3VL). The main purpose is to show via experiments on several public data 1) 3VL is more accurate than binary logic for interpreting a trained MLP, 2) the MLP can be more interpretable if we use structural learning with forgetting, and 3) the performance of the discretized MLPs can be improved through retraining. Based on the results obtained here, we will point out some important topics for further study.