Visualization of hidden node activity in neural networks.

Wlodzislaw Duch,
Department of Informatics, Nicolaus Copernicus University,
Grudziadzka 5, 87-100 Torun, Poland,
and School of Computer Engineering, Nanyang Technological University, Singapore.

Final version consists of two parts, therefore two papers are attached here.
Published in Lecture Notes on AI, Lecture Notes in AI Vol. 3070 (2004) 38-43; 44-49
Presented at the International Conference on Artificial Inteligence and Soft Computing (ICAISC), Zakopane 2004.

Abstract: Visualization of hidden node activity in neural networks: I. Visualization methods.
Quality of neural network mappings may be evaluated by visual inspection of hidden and output node activities for the training dataset. This paper discusses how to visualize such multidimensional data, introducing a new projection on a lattice of hypercube nodes. It also discusses what type of information one may expect from visualization of the activity of hidden and output layers. Detailed analysis of the activity of RBF hidden nodes using this type of visualization is presented in the companion paper.

Abstract: Visualization of hidden node activity in neural networks: II. Application to RBF networks.
Scatterograms of images of training vectors in the hidden space help to evaluate the quality of neural network mappings and understand internal representations created by the hidden layers. Visualization of these representations leads to interesting conclusions about optimal architectures and training of such networks. Depending on network parameters only some parts of the unit hypercube -- called here admissible spaces -- may be reached. The usefulness of visualization techniques is illustrated on parity problems solved with RBF networks.

Part I, paper in PDF, 114 KB

Part II, paper in PDF, 430 KB

BACK to the publications of W. Duch.
BACK to the on-line publications of the Department of Informatics, NCU.