Artificial neural networks are computer software or hardware models inspired by the structure and behavior of neurons in the human nervous system. As a powerful learning tool, increasingly neural networks have been adopted by many large-scale information processing applications but there is no a set of well defined criteria for choosing a neural network. The user mostly treats a neural network as a black box and cannot explain how learning from input data was done nor how performance can be consistently ensured. We have experimented with several information visualization designs aiming to open the black box to possibly uncover underlying dependencies between the input data and the output data of a neural network. In this paper, we present our designs and show that the visualizations not only help us design more efficient neural networks, but also assist us in the process of using neural networks for problem solving such as performing a classification task.

In the traditional volume visualization paradigm, the user specifies a transfer function that assigns each scalar value to a color and opacity by defining an opacity and a color map function. The transfer function has two limitations. First, the user must define curves based on histogram and value rather than seeing and working with the volume itself. Second, the transfer function is inflexible in classifying regions of interest, where values at a voxel such as intensity and gradient are used to differentiate material, not taking into account additional properties such as texture and position. We describe an intuitive user interface for specifying the classification functions that consists of the users painting directly on sample slices of the volume. These painted regions are used to automatically define high-dimensional classification functions that can be implemented in hardware for interactive rendering. The classification of the volume is iteratively improved as the user paints samples, allowing intuitive and efficient viewing of materials of interest.

In volume data visualization, the classification step is used to determine voxel visibility and is usually carried out through the interactive editing of a transfer function that defines a mapping between voxel value and color/opacity. This approach is limited by the difficulties in working effectively in the transfer function space beyond two dimensions. We present a new approach to the volume classification problem which couples machine learning and a painting metaphor to allow more sophisticated classification in an intuitive manner. The user works in the volume data space by directly painting on sample slices of the volume and the painted voxels are used in an iterative training process. The trained system can then classify the entire volume. Both classification and rendering can be hardware accelerated, providing immediate visual feedback as painting progresses. Such an intelligent system approach enables the user to perform classification in a much higher dimensional space without explicitly specifying the mapping for every dimension used. Furthermore, the trained system for one data set may be reused to classify other data sets with similar characteristics.

In volume visualization, users typically specify transfer functions to classify the data and assign visual attributes to each material class. Higher-dimensional classi�cation makes it easier to differentiate material classes since more data properties are considered. One of the dif�culties in using higher-dimensional classi�cation is the absence of appropriate user interfaces. We introduce an intuitive user interface that allows the user to work in the cluster space, which shows the material classes with a set of material widgets, rather than work in the transfer function space. This interface not only provides the user the capability to specify arbitrary-dimensional transfer functions, but also allows the user to operate directly on the classi�cation and visualization results.