Fine tuning of adaptive learning of deep belief network for misclassification and its knowledge acquisition
Shin Kamada, Takumi Ichimura · International Journal of Computational Intelligence Studies · 2017
We have proposed an adaptive structure learning of deep belief network (DBN) that can determine the suitable number of hidden layers and hidden neurons of restricted Boltzmann machines (RBMs). The method shows high classification performance to the big data benchmark test. However, the method could not classify the unknown pattern correctly, since an input data with ambiguous patterns leads the classification to the wrong judgment. In such a case, a fine-tuning method that patches a part of network signal flow based on the knowledge will be a helpful method even in terms of both the improvement of classification capability and the reduction of computational cost by learning again. In this paper, network signal patterns which lead the given misclassified patterns were visualised for knowledge acquisition. By fine-tuning the trained network using the acquired knowledge, the classification capability can achieve great success.