An Approach to Searching for Two-Dimensional Cellular Automata for Recognition of Handwritten Digits

Abstract

One of the contexts in which cellular automata have clearly demonstrated their effectiveness has been in problems involving strong and explicit spatial constraints, as happens in pattern formation and growth. By analogy, attempts to use cellular automata in pattern recognition have also been used in the literature and some progress has been made. However, in general, they still represent more of an unfulfilled promise, due to the lack of a recognition model which cellular automata would naturally fit in, the lack of effective ways to implement it, and the lack of generality of the available approaches. Here, experimental results are reported in the direction of using cellular automata in the task of handwritten digit recognition, in which an evolutionary algorithm searches for two-dimensional cellular automata rules that would transform a given digit image into a match, as close as possible, to a prototype image of that family, so that, the closer the match, the better the recognition of the input image. Although the results reported might still fall shorter than consolidated commercial techniques for the task, the approach presented is quite attractive in terms of the efficacy level it allowed to achieve, and because of its simplicity, which suggests a potential generality from the perspective of its use in other domains.