Segmentation

This is unique to strong TC. Clearly, not every possible sequence of classes can be learned. Consider the speech recognition task, this would be like trying to learn how to recognise every possible sentence. Any training set is unlikely to contain even a small percentage of the possible sentences; how would it be possible to make a system that classifies all possible sentences?

The most obvious solution is to break the training sentences into individual words, learn to recognise them, and then somehow combine these individual word recognisers into something that can recognise whole sentences. However, this introduces other problems. If all that is given is a training stream and the corresponding class sequence, when does one class in the class sequence end and the next begin? How are ``transition periods'' from one class to the next handled? How does one cope with the problem that classes in the sequence are not independent? If the stream is segmented incorrectly then any learner will then get the wrong data: one class will get a part of a stream that belongs to the next class, and the other will get less frames then it should get.