The unit structure

The units together with the links determine the network topology. Connected to each unit is a linked list of links that specifies the units or external input terminals which are the inputs to the unit. A second linked list of links determines the units to which the unit output is transferred. Thus, given a unit in a network all inputs and outputs can be found immediately.

Besides pointers to the links, the unit contains a unique identification number (which can be used for searching), a layer identification number (if appropriate) and control information. This control information specifies the type of unit (e.g. input, hidden or output, see UNIT-flags) and allows the possibility to disable a complete unit. The input/output index specifies the position of the unit in the input layer or output layer in case the unit flags are INUNIT or OUTUNIT, see (see NET-flags).

  
Figure 3.2: The unit data structure.

A special weight, called the theta weight (i.e. bias weight) is stored in the unit. Its structure is equal to all other weights and is used in some network paradigms. A pointer to the unit activation function and unit transfer function allows different types of units to be mixed in the same network. The standard supported activation functions are the weighted sum of inputs and weights, the Euclidean or squared Euclidean distance of inputs and weights. The user can easily define his own activation function and put the pointer in the unit to point to the new function, see UNIT.

The transfer functions are implemented in a similar way. Standard transfer functions are linear (see TransLinear), sigmoidal (TransSigmoid), tanh (TransTanH) and Gaussian (TransGauss). The user can point to a user defined transfer function if needed. For learning methods that require the derivative of the transfer function or even a second derivative, a pointer to such a derivative can be stored in the unit structure too, see UNIT.

A counter of the number of input links and the number of output links, although redundant, is added for consistency checks. General purpose data space enables the user to add his own information to this structure.