Classification Models and "Strength" Values

Neuralware's Predict classification models produce predicted outputs in the same format as the targets in your original data. For example, model outputs might be in the form of 0/1, yes/no, red/green/blue, class1/class2/class3/class4 or some other format. While that kind of output is certainly useful, it is often desirable to also know the strength of each predicted class assignment. Information can be gained, for example, by looking at two records that both predict "yes". One record might predict 0.60 as "yes" and another predict 0.99 as "yes"; it is important to know that you can have more confidence in the second data record.

Predict allows you to control output options via the Predict menu. Simply choose Predict/Run, put a check mark in the check box Internal Units (Network output not converted), and continue the task by clicking the Run button. The predicted outputs will then include continuous values. For example, a four- class output, e.g., {red, orange, yellow, blue} would generate four values, e.g., {0.01, 0.42, 0.23, 0.34}, and the highest value would be chosen as the winning class. For this particular record, orange (0.42) would be the surviving predicted output.

We're throwing down the gauntlet! Send us your best model and the data you used to create it. Using NeuralSight, we will attempt to beat your results, proving NeuralSight's value. NeuralSight works like a team of 100 modelers staying up all night testing different neural network models and finding the best ones.

We can only accept a limited number of challenges, and we'll have some specific requirements about the data you give us. If you would like take the NeuralSight challenge, simply send email to Bob Everly, and he'll set it up. Good luck!