The general case of using patterns with a neural network is an exact fit of the patterns onto the network. The set of activations of all input units is called input pattern, the set of activations of all output units is called output pattern. The input pattern and its corresponding output pattern is simply called a pattern. This definition implies that all patterns for a particular network have the same size. These patterns will be called regular or fixed sized.
SNNS also offers another, much more flexible type of pattern. These patterns will be called variable sized. Here, the patterns are usually larger than the input/output layers of the network. To train and recall these patterns small portions (subsequently called subpatterns) are systematically cut out from the large pattern and propagated through the net, one at a time. Only the smaller subpatterns do have to have the fixed size fitting the network. The pattern itself may have an arbitrary size and different patterns within one pattern set may have differing sizes. The number of variable dimensions is also variable. Example applications for one and two variable dimensions include time series patterns for TDNNs and picture patterns.
Both of these types of patterns are loaded into SNNS from the same
kind of pattern file. For a detailed description of the structure of
this file see sections 
and . The grammar is
given in appendix