Error Files

Error files (generated with the -E command line parameter) can sometimes provide insight into why the network failed to learn the target concept exactly. Error files can only be created when testing using labeled examples and using the default output mode (-o accuracy). The created error file will contain a list of examples for which SNoW made an incorrect prediction. Each example's listing shows the example number and each target's prediction confidence and activation.

For example, in the following error file, we see there were mistakes on the and examples in the input file. We also see the activations calculated by each target. The first activation is the prediction confidence (see Section 4.3.5 for more details on prediction confidence). The second is the target's raw activation (in other words, it is simply the dot product of the target's weights and the example^6.2). The correct target's output is marked with an asterisk. Note that this output is the same as that generated by the -o allactivations output mode.

Ex: 3 Prediction: 1 Label: 0
1: 0.1824 4.5
0: 0.1192 4*

Ex: 4 Prediction: 1 Label: 0
1: 0.19193 4.5625
0: 0.18243 4.5*

Footnotes

... example^6.2

In the case of clouds containing more than one target node, this activation is the unweighted summation over all target nodes that learn the class label of those dot products.