Much of the current interest in neural networks can be traced to the introduction a few years ago of effective learning algorithms for these systems ([Denker:86a], [Parker:82a], [Rumelhart:86a]). In [Rumelhart:86a] Chapter 8, it was shown that for some problems using multi-layer perceptrons (MLP), back-propagation was capable of finding a solution very reliably and quickly. Back-propagation has been applied to a number of realistic and complex problems [Sejnowski:87a], [Denker:87a]. The work of this section is described in [Felten:90a].
Real-world problems are inherently structured, so methods incorporating this structure will be more effective than techniques applicable to the general case. In practice, it is very important to use whatever knowledge one has about the form of possible solutions in order to restrict the search space. For multilayer perceptrons, this translates into constraining the weights or modifying the learning algorithm so as to embody the topology, geometry, and symmetries of the problem.
Here, we are interested in determining how automatic learning can be improved by following the above suggestion of restricting the search space of the weights. To avoid high-level cognition requirements, we consider the problem of classifying hand-printed upper-case Roman characters. This is a specific pattern-recognition problem, and has been addressed by methods other than neural networks. Generally the recognition is separated into two tasks: the first one is a pre-processing of the image using translation, dilation, rotations, and so on, to bring it to a standardized form; in the second, this preprocessed image is compared to a set of templates and a probability is assigned to each character or each category of the classification. If all but one of the probabilities are close to zero, one has a high confidence level in the identification. This second task is the more difficult one, and the performance achieved depends on the quality of the matching algorithm. Our focus is to study how well an MLP can learn a satisfactory matching to templates, a task one believes the network should be good at.
In regard to the task of preprocessing, MLPs have been shown capable [Rumelhart:86a] Chapter 8 of performing translations at least in part, but it is simpler to implement this first step using standard methods. This combination of traditional methods and neural network matching can give us the best of both worlds. In what follows, we suggest and test a learning procedure which preserves the geometry of the two-dimensional image from one length scale transformation to the next, and embodies the difference between coarse and fine scale features.