Accountability of neural networks trained with “real world” data

In the real world neural networks have to be accountable. It is not sufficient to produce output data from a network without any information as to the quality of that data. The quality of the output data from any given neural network for a given generalisation point is dependent on the following: (1) the representative nature of the original training data in relation to the scope of the problem domain; (2) the properties of the network training algorithm, and degree of training; (3) whether the generalisation point lies within the domain of validity of the network; (4) the density of training data in the region of the generalisation point; and (5) the quality of the training data in the region of the generalisation point. The first of these is outside the scope of this paper. The second point is a problem that should be well known to those working in the field and it is not our intention to cover general training and validation principles in this paper. We address the other three points raised above. Firstly we introduce each of the points in greater detail, outlining the pitfalls open to the unwary. Next, for the domain of validity and the data density problems, we review some of the methods available to the practising connectionist; finally, for the problem of local data quality, we outline current research being carried out by the authors into this area