Neural networks using modified initial connection strengths by the importance of feature elements

Abstract

In this paper, a neural network method is applied to extract one cycle of golf swing from a continuous weight-shift signal. Weight-shift in golf swing means the continuous change of weights loaded on the left and right foot of the golfer. We defined eight input features which are stable to classify various shapes of swing patterns. The adopted network is a three-layered error backpropagation model. According to experimental results, identifying success rate is 97.75% using 8 input, 10 hidden and 2 output nodes. We performed experiment by changing the initial connection strengths according to a importance scale. Under ten random seeds, the learning speed and recognition rate is shown to improve when the initial connection strengths are changed by the importance scale.