BP neural network is an important part of neural network, and is broadly used in many areas. For BP neural network it's essential to learn a lot of training samples. However, the dataset of the training samples is so large and multi-dimensional that traditional techniques can't convey them effectively in a relative small display area. To address this problem, a parallel coordinates-based BP network visualization technique is proposed. In this approach, the parallel coordinates is constructed according to the model of BP neural network. In particular, clusters, colors, size, density and other features are used to improve the parallel coordinates to create clear intelligible visualization. At last, an experiment is designed to argue that parallel coordinates-based BP network visualization is feasible and effective, and can reveal the relations and rules hidden in the training samples in the process of the training.
Unfortunately, ACM prohibits us from displaying non-influential references for this paper.
To see the full reference list, please visit http://dl.acm.org/citation.cfm?id=1865852.