1 of 14
Every time you begin Tiberius there are several options... 2 of 14
When loading in data you can select Excel, Access or any other type of ODBC data source. 3 of 14
Select the appropriate table, query or worksheet and the available variables will be shown.
Choose what you want as inputs (the independent variables) and what is to be the output (the dependent variable). 4 of 14
All new networks start with one linear hidden neuron, which is equivalent to linear regression.
Extra neurons can be added with the click of a mouse. 5 of 14
With linear regression, the function y=x² cannot be solved. 6 of 14
One non-linear neuron cannot model the data either, but it is an improvement... 7 of 14
Two non-linear neurons are required to model this data.
An anomoly in the data can be spotted. 8 of 14
With Tiberius you can decompose the neural model to see the contribution of individual neurons.
Here it can be seen how two neurons combine to model the data. 9 of 14
A plot of the model error highlights an anomoly in the data. 10 of 14
Use the mouse to zoom in on the area of interest, and remove the 'odd' data point from the training data. 11 of 14
Cleaning up the data will improve the neural model. 12 of 14
Regions of the data can be selected for a training and test set. 13 of 14
It can be seen here how just training on the centre region gives a good model for this data, but outside the training range the model is not as good.
Like all other modelling techniques, neural network models should not be used to extrapolate into the unknown. 14 of 14
The error plot shows how the model error is greater in the regions where the network is extrapolating into the unknown (the input values are outside the training range of experience).
Tiberius has tools that will prevent extrapolation into the unknown.