SNoW may be run in five major modes: Training, Testing, Interactive, Evaluation, and Server.
Training consists of presenting labeled examples (see Chapter 6 for a description of file formats) to the network and learning a weight vector representation for each target concept. At the completion of training, the resultant network, which can be used to make predictions on previously unseen examples, is written to the network file. Thus, training must always be performed before testing or evaluations can be performed.
Testing consists of presenting examples to the system and predicting a label. This is done by evaluating the representation learned for each of the targets - computing the target's prediction confidence given the new example - and choosing the target with the highest prediction confidence. The results of testing can be output in a number of ways.
Interactive mode allows the user to (1) evaluate specific examples, and (2) control the training process by specifying the targets to be promoted or demoted on an example by example basis. In this mode, the updating decision is usually made based on previous results of evaluation. For example, one may want to evaluate one example first, look at the activation value of each class, and then decide which classes to promote and which to demote. Therefore, to make the interactive mode truly interactive, the input file shouldn't be a regular file, but STDIN or a pipe instead.
Evaluation is used to make a prediction based on a labeled or unlabeled example supplied on the command line.
Server mode is a limited version of Testing mode in which after loading the network, SNoW waits for examples to be sent over a user specified port. SNoW then returns its output over the same port. This mode can make testing much more efficient when the network is large and there are many example files that are best or necessarily kept separate when tested. Time will not be spent reloading the network for each file.
Several other modes of operation initiated with optional command line parameters are supported, including, but not limited to, incremental learning (see option -i), regularization (Section 4.3.2), and function approximation (Section 4.3.3).
A recommendation: SNoW has several options and input parameters that are required and others that are commonly used. We recommend saving commands and options in a simple shell script that records the typical way you use SNoW, and running this script. This also allows you to automate parameter tuning in a simple way. In SNoW's distribution, you will find a booleanexample/ directory corresponding to the discussion in Section 6.1.3 and a tutorial/ directory corresponding to the discussion in Chapter 7, both of which contain examples of SNoW's usage in easy to use shell scripts.