Inducing High Performance and Compact Neural Networks Based on Decision Boundary Making
Yuya Kaneda, Qiangfu Zhao · IEEJ Transactions on Electronics Information and Systems · 2014
In recent years, portable computing devices (PCDs) are becoming very popular. To improve the quality of service (QoS) for each individual user, it is necessary to develop application programs that can be aware of the user intention, preference, situation, etc., so that proper services can be recommended at proper timing. We call these kinds of programs awareness agents (A-agents). To satisfy various needs of a user, many A-agents should work together in one PCD. Since the computing resource in a PCD is limited, it is necessary to reduce the implementation costs of the A-agents while preserving their performance. For this purpose, we propose two decision boundary making (DBM) algorithms in this paper. The basic idea of DBM is to generate new training data using given ones to fit the decision boundary (DB) of the given problem, and induce small neural networks (NNs) using the new data. Both algorithms proposed here are simplified versions of the decision boundary learning (DBL) algorithm proposed by us earlier. Using the new algorithms, the cost for data generation can be greatly reduced. Experimental results show that if the new data are generated properly in positions close to the DB, the induced small NNs can perform even better than support vector machines, which are known as the state-of-the-art learning models.