For practical purposes, it would be useful if time to train from the data is small. There is also an issue of whether the learner could train incrementally. In many domains, we would also like ``on-line'' testing to be possible. This is where we classify signals as they arrive. For example, one possible example would be to collect sensor data from a robot, train from the data in batch (off-line) and then to install the classifier produced by training on the robot once again, such that as soon as sensor readings arrive, and as soon as a particular class is observed, a classification is produced.