At present, only two network architectures are supported: A feed-forward network with 2-4 layers, or a simple recurrent network with a single hidden unit layer.
The network can be configured using ‘Network->Configure Network’, which brings up a tabbed pane. Choose either ‘FeedForward Net’ or ‘SRN’ (Simple Recurrent Network). For each layer except the output layer, you may opt to add a bias unit. This acts as a bias for the next layer up, so a bias on the input layer adds a constant term to the sum calculated by units in the first hidden unit layer, etc.
If an SRN is configured, the hidden layer is shown with a blue enclosing ring, as a visual reminder that the units in this layer are recurrently connected.
Network configurations can be saved and loaded through the ‘Network’ menu. An example of the text form for the above network is shown below:
[Type] Type = SRN [Network] Layers = 3 Input units = 1 Input bias = true Layer1 units = 5 Layer1 bias = true Output units = 1
Here is a 4-layer feed forward network and its associated configuration file:
And the configuration file:
[Type] Type = FeedForward [Network] Layers = 4 Input units = 2 Input bias = true Layer1 units = 5 Layer1 bias = true Layer2 units = 4 Layer2 bias = false Output units = 1