Data:
each row in the matrix is a house, characterized by a series of parameters like: median value in dollars,
number of rooms, age, crime rate in town, nitric oxides concentration, accessibility to radial highways, etc.
Objectives of neural network modelling:
- To build a model associating selected inputs (for example number of rooms, age, crime rate) to
one output, for example the house value. The model can then be used to predict house values for
new houses.
- To estimate the prediction expected prediction error of the neural network model.
- To identify which inputs are more relevant to predict the output.
- To quickly analyze how the output depends on selected inputs by using
the Grapheur Output Sweeper tool.
Grapheur sample visualization: neural network training and testing
The form of the model, the input(s) and the output to be predicted are selected.
Then a training session is started and the obtained training and testing error
during training are immediately visualized. One can easily experiment with different
architectures and save the best predictor when finished. In the image below one sees
a steady decrease of the training error (red curve), while the test error first
decreases and then reaches a plateau. Continuing training after reaching this plateau
can in some cases lead to over-training (memorization of the training data with
poor generalization, see manual for additional details).
Grapheur sample visualization: Grapheur Output Sweeper
After a neural network model is built, the Grapheur Output Sweeper tool can be
used for visualizing the output when the various inputs are changing.
Two input variables are fixed as X and Y coordinates, the output is shown
with different colors. The value of the other input variables can be fixed by
the corresponding slider. A specified input variable can be selected for
an automated sweep over all possible values, showing a moving slice of
the input space.
