An Incremental Learning Network Model Based on Random Sample Distribution Fitting

The training of the classification network has tough requirements of data distribution. The more the training data did not consistent with the distribution of the real target function, the higher error rate the network will produce. In the context of incremental learning, the data distribution of the subsequent training tasks may not consistent with the data distribution of the previous tasks. To handle this problem, lots of learning methods were introduced, most of these methods are complicated and heavy computing. In this paper, a novel method which is faster and simpler is proposed to uniform subsequent training data. Artificial training samples are produced from random inputs in current trained network. In subsequent task, these artificial samples were mixed with new real data as training data. The experiments with proper parameters show that new features from new real data can be learned as well as the old features are not forgot catastrophically.