Aspects of Application of Neural Recognition to Digital Editions

Artificial neuronal networks (ANN) are widely used in software systems which require to solve problems without a traditional algorithmic approach, like in character recognition: ANN learn by example, so that they require a consistent and well-chosen set of samples to be trained to recognize their patterns. The network is thaught to react with high activity in some of its output neurons whenever an input sample belonging to a specified class (e.g. a letter shape) is presented to it, and has the capability of assessing the similarity of samples never encountered before to any of these models. Typical OCR applications thus require a significant amount of preprocessing for such samples, like resizing images and remove all the ‘noise’ data letting the letter countours emerge clearly from the background. Further, usually a huge number of samples is required to effectively train a network to recognize a character against all the others. This may represent an issue for paleographical applications, because of the relatively low quantity and high complexity of digital samples available, and poses even more problems when our aim is detecting subtle differences (e.g. the special shape of a specific letter from a well-defined period and scriptorium). It would be probably wiser for scholars to define some guidelines for extracting from samples the features defined as most relevant according to their purposes, and let the network to deal with just a subset of the overwhelming amount of detailed nuances available. ANN are no magic, and it is always the careful judgement of scholars to provide a theorical foundation for any computer-based tool they might want to use to help them solve their problems: we can easily illustrate this point with samples drawn from any other application of IT to humanities. Just as we can expect no magic in detecting alliterations in a text if we just feed a system with a bunch of letters, we can no more claim that a neural recognition system might be able to perform well with a relatively small sample where each shape is fed as it is, without instructing the system about the features scholars define as relevant. Even before ANN implementations, it is right this theorical background which must be put into test when planning such systems.