Evenet 2000: Designing and Training Arbitrary Neural Networks in Java
In this paper, Evenet-2000, a Java-Based neural network toolkit is presented. It is based on the representation of an arbitrary neural network as a block diagram (these blocks are, for example, summing junctions or branch points) with a set of simple manipulation rules. With this toolkit, users can easily design and train any arbitrary neural network, even time-dependent ones, avoiding the complicated calculations that the means of establishing the gradient algorithm requires when a new network architecture is designed. Evenet-2000 consists of three parts: a calculation library, a user-friendly interface and a graphic network editor with all the Java advantages: encapsulation, inheritance, powerful libraries...
KeywordsNeural Network Gradient Algorithm Graphic Editor Signal Flow Graph Calculation Library
Unable to display preview. Download preview PDF.
- 3.Piñeiro J.D., Marichal R.L., Moreno L., Sigut J., Estevez I, Aguilar R., Sanchez J.L. and Merino J. “Evoked potential feature detection with recurrent dynamic neural networks.” Proceedings in International ICSC/IFAC Symposium on Neural Computation NC’98. Vienna Austria, 1998Google Scholar
- 4.Campolucci P., Marchegiani A., Uncini A., Piazza F., “Signal-Flow-Graph Derivation of On-line Gradient Learning Algorithms” IEEE International Conference on Neural Networks, Houston (USA), June 1997.Google Scholar
- 5.Osowski S., Herault J. “Signal Flow Graphs as an Efficient Tool for Gradient and Exact Hessian Determination”. Complex Systems, Vol 9, 1995.Google Scholar