Induction of High Performance Neural Networks Based on Decision Boundary Making

Smartphone, in recent year, becomes popular and has been widely applied by users. In order to meet different needs from users, embedding "awareness" providing supports by understanding onto smartphone devices is necessry. Due to the limitations (e.g. computing resources, etc.) on smartphone, methods that is light but with high performance are strongly expected. In this study, the concept of awareness agent (A-agent) is proposed for the purpose. For this purpose, we have proposed decision boundary learning (DBL) based on particle swarm optimization (PSO). Results show that this method can yield compact neural network (NNs) agents that are comparable in performance to support vector machines (SVMs). However, the computational cost of PSO is high, and the method cannot be used in smartphone environments. To reduce the computational cost, we propose a simple method called decision boundary making (DBM). The basic idea of DBM is to generate new training data around the support vectors of an SVM, add them to the training set, and then induce an NN agent. We conducted experiments using several public databases, and experimental results show that the proposed DBM is comparable to DBL in performance, and the computational cost can be greatly reduced.