Neural Network to Extract Frequent Sub-sequences from Symbol Sequences with Fluctuation in Appearance Interval

Abstract

This study aimed to extract words as frequent sub-sequences from voice data given in streaming format. In the previous method, in the case of the symbol sequence which has fluctuation in the appearance interval of the symbol like voice, the network cannot be learned efficiently. In this paper, we propose an efficient learning method and aim to extract frequent subsequences faster than the previous method. In the previous method, a unit called LIF (Leaky Integrate and Fire) model is used. We proposed delaying the firing timing of this unit to adjust the connection weight more efficiently by STDP learning rule. In order to confirm the effectiveness, the length was limited to 2 symbol lengths for simplification, we measured the time which is taken to extract frequent sub-sequences by the previous method and the proposed method. By delaying the firing timing of this unit, we confirmed the spiking neural network can extract frequent sub-sequences faster than the previous method.