Kohonen Node Expert Options
For those with detailed knowledge of Kohonen networks, expert options allow you to fine-tune the training process. To access expert options, set the Mode to Expert on the Expert tab.
Width and Length. Specify the size (width and length) of the two-dimensional output map as number of output units along each dimension.
Learning rate decay. Select either linear or exponential learning rate decay. The learning rate is a weighting factor that decreases over time, such that the network starts off encoding large-scale features of the data and gradually focuses on more fine-level detail.
Phase 1 and Phase 2. Kohonen net training is split into two phases. Phase 1 is a rough estimation phase, used to capture the gross patterns in the data. Phase 2 is a tuning phase, used to adjust the map to model the finer features of the data. For each phase, there are three parameters:
- Neighborhood. Sets the starting size (radius) of the neighborhood. This determines the number of "nearby" units that get updated along with the winning unit during training. During phase 1, the neighborhood size starts at Phase 1 Neighborhood and decreases to (Phase 2 Neighborhood + 1). During phase 2, neighborhood size starts at Phase 2 Neighborhood and decreases to 1.0. Phase 1 Neighborhood should be larger than Phase 2 Neighborhood.
- Initial Eta. Sets the starting value for learning rate eta. During phase 1, eta starts at Phase 1 Initial Eta and decreases to Phase 2 Initial Eta. During phase 2, eta starts at Phase 2 Initial Eta and decreases to 0. Phase 1 Initial Eta should be larger than Phase 2 Initial Eta.
- Cycles. Sets the number of cycles for each phase of training. Each phase continues for the specified number of passes through the data.