New Features of SNNSv4.0
Version 4.0 of SNNS features the following improvements and extensions
over the earlier version 3.3:
- A new distributed version of the kernel. SNNS can now be spread
out in a workstation cluster for faster learning.
- Improved version of the C-code generator snns2c
- Validation sets now can check on the performance of the network
during training.
- New error function in tool analyze that can handle
single-output-unit-networks.
- New statistic tool that can predict the generalization
capabilities of the network.
- Reorganization of the manager panel with new selection mechanism
for the major SNNS windows.
- The remote panel was changed significantly and renamed
`control panel'.
- Selection possibility in the graph panel between plot of SSE,
MSE, and SSE/output.
- New learning algorithm RBF\_DDA
- New learning algorithm simulated annealing
- New learning algorithm Monte Carlo.
- New learning algorithm Pruned-Cascade-Correlation
- Several new initialization algorithms for Kohonen and
Counterpropagation networks, since the two old ones had serious
flaws.
- Key codes to bring up all the main panels of SNNS (e.g Alt-c for the
control panel).
- Support for NeXT systems.
- More information printed to the shell during training and with
the INFO button.
- The SOURCE and TARGET part of the info panel have been swapped
for a more intuitive setup of the information.
- Extensive debugging (as usual).
Last modified: mache@rembrandt Thu May 18 15:11:32 MET DST 1995