next up previous contents
Next: Discussion Up: Experiments Previous: Performance with the speech

Performance with the Boston data

The experiments with Boston data were run four times with random division of data into training and test sets. Mean square reconstruction errors were scaled such that the SOM errors were 100 in each case. The NFA errors were 118 with standard deviation of 17 and the FA errors were 151 with standard deviation of 22.

This data set is clustered, because of the town structure, and the dimensionality of the data manifold is not too large for the SOM to handle. Therefore it is not very surprising that it made better reconstructions than NFA. Nonlinearities were crucial in the data set, since FA was inferior to the nonlinear methods. NFA was run with 20 hidden neurons and from 1 to 9 factors. Best number of them varied from 5 to 9. The best number model vectors in the SOM varied from 500 to 1800, which is far greater than the number of data points.

Tapani Raiko