A STRATEGY FOR CREATING INITIAL DATA ON ACTIVE LEARNING OF MULTI-LAYER PERCEPTRON

摘要

Many active learnings in the training of a partially trained Multi-Layer Perceptron (MLP) have been proposed. We note any active learning performance depends on initial training data. The initial training data plays an important role for active learning performance, because any active learning algorithm generates additional training data that is useful for improving the classification accuracy, based on initial training data. Most of conventional methods have generated initial data at random using a pseudo-random number. However, in practical case, we can not prepare enough data by the limit of time and cost. Therefore, the bias of initial training data becomes critical, especially in the case of input space dimension to be large. In this paper, we propose a strategy by the use of low-discrepancy sequence for creating more uniform initial data than pseudo-random numbers. For the classification problem of MLP, we analyze the experimental performances of network inversion algorithm which use a pseudo-random number and a low-discrepancy sequence as initial training data. In experimental results, we found low-discrepancy sequences give a good strategy to create initial training data. Finally, we also discuss some advantages and disadvantages of low discrepancy sequences as initial training data.