I/O Layers

The I/O (Input / Output) layers represent interfaces between the processing layers of a neural network and the external environment, providing the net with the data needed for processing and/or training.

File Input

The file input layer allows data in a file to be applied to a network for processing. Data for processing is expected as a number of rows of semicolon-separated columns of values. For example, the following is a set of three rows of four columns:

0.2;0.5;0.6;0.4
0.3;-0.35;0.23;0.29
0.7;0.99;0.56;0.4

Each value in a row will be made available as an output of the file layer, and the rows will be processed sequentially by successive processing steps of the network.
As some files may contain information additional to the required data, the parameters firstRow, lastRow, firstCol and lastCol may be used to define the range of useable data. The filename parameter specifies the file that is to be read from.

URL Input

The URL input layer allows data from a remote location to be applied to a network for processing. The allowed protocols are http and ftp. The data format is the same as for the FileInput layer.

Excel Input

The Excel Input layer permits data from an Excel file to be applied to a neural network for processing. Its ‘sheet’ parameter allows the name of the sheet to be chosen from which the input data is read.

Switch Input

The switch input allows the choice of which input component is to be connected to the neural network, choosing between all the input components attached to it. The user, after having attached several input components to its input, can set the ‘active input’ parameter with the name of the chosen component that is to be connected to the net. The ‘default input’ parameter must be filled with the name of the default component (the one activated when the user selects the ‘Control->Reset Input Streams’ menu item).

The switch input component, along with the output switch layer, permits dynamic changing of the architecture of the neural network, changing the input and/or output data layers attached to the neural network at any time. This is useful to switch the input source, for instance, between the file containing the training data set and the file containing the validation data set to test the training of the neural network, as depicted in the following screen shot:

input switch

Learning Switch

The learning switch is a special implementation of the Switch Input component and can be used to attach both a training data set and a validation data set to the neural net. In this way the user can test the generalization property of a neural network using a different data set to the one used during the training phase.

The training input data set can be attached by dragging an arrow from the input component to the learning switch, while the validation input data set can be attached simply by dragging an arrow from the red square on top of the learning switch to the input component containing the validation data set. To switch between them, simply change the value of the 'validation' parameter shown in the Control Panel.

The following figure depicts the use of this component:

learning switch


Warning: Because a validation data set will also be required for the desired data, this component must be inserted both before the input layer of the neural network and between the Teacher layer and the desired input data sets.

File Output

The file output layer is used to convert the results of a processing layer to a text file. The filename parameter specifies the file that the results are to be written to. Results are written in the same semicolon-separated form as file input layers.

Excel Output

The Excel output layer is used to write the results of a processing layer to an Excel formatted file. The filename parameter specifies the file that the results are to be written to. The ‘sheet’ parameter allows the name of the sheet to be chosen, to which the input data is to be written.

Switch Output

The switch output permits the choice of which output component is to be connected to the neural network, choosing between all the output attached components. The user, after having attached several components to its output, can set the ‘active output’ parameter with the name of the chosen component that is to be connected to the net. The ‘default output’ parameter must be filled with the name of the default component (which one activated when the user selects the ‘Control->Reset Input Streams’ menu item).

Teacher

The Teacher layer allows the training of a neural net, using the back-propagation learning algorithm. It calculates the difference between the actual output of the net and the expected value from an input source. It provides this difference to the output layer for the training.

To train a net, add a Teacher component, connecting it to the output layer of the net, and then connect an input layer component to it, linked to a source containing the expected data (see figure).