The links make up the connections of the units. The direction of the internal pointers of the link data structure is from the target unit (destination unit) to the source unit (origin unit). Note that this is reversed from the direction of the links in the neural network (which run from source to target unit). This approach allows the efficient storage of the network structure with a minimum of memory. However, the memory needed to store big nets is still the limiting factor. For a net with 10.000 units with a connectivity of 20%, 20 million connections have to be stored, requiring 230 megabyte of memory. In comparison, the fraction of the memory used to store all other components is negligible.
The structure of a link is quite similar to that of a site:
Like sites, the links are grouped in linked arrays which are handled by the SNNS memory management routines.
