| Popis: |
Input and training signals are presented to recurrent artificial neural networks in an inherently sequential manner. However, of the two broad categories of practical problems to which recurrent neural networks are applied, signal prediction/reproduction and supervised signal classification, the information necessary to construct a training signal that can be presented over the full length of the input signal is usually only present in the first category. It is this absence of information in the second group that gives rise to differing problems, techniques and nomenclature in what could be otherwise be considered a subset of the signal prediction problem. The only information usually available for use as a training signal with classification tasks is the desired final classification of each training example. Normal practice is to present this piece of information, as a training signal of length one, concurrently with the last piece of input data, ignoring the possibility of providing a training signal prior to that point. The sequential nature of the training signal for classification tasks has been, to all intents and purposes, forgotten. This work explores the deficiency by examining the nature of suitable values, whether or not they can be constructed from the available information, and whether or not standard neural network training algorithms can be (or need to be) modified to make use of such values. It finds that methods do exist and develops a family of such methods, signal melding, which demonstrates the usefulness and possibilities inherent in filling in the missing values in the training signal. The signal melding technique makes use of the dynamic information seen at a network's outputs during the presentation of input. As a benefit, this information offers some important clues about the diversity of examples within an apparently homogeneous class. To make full use of these values, a modification to back propagation through time is developed, attentive back propagation through time, which shows clear advantages over the standard algorithm when using training signals over the full length of the input signal. Simulations taking over three years of run-time are presented which show that on simple classification tasks, when signal melding and attentive back propagation through time are combined, performance levels at least comparable with existing methods are achieved, together with much improved stability. |
