Handwritten Digit Recognition Based on Classification Functions

As a model of a machine learning, an incompletely specified classification function is used. As a benchmark problem, data for handwritten digits with 28 x 28 images were used. This data was converted into one with 14 x 14 = 196 pixels using a space filter. Also, the value of each pixel was binarized. With this operation, the original data was converted into a 196-variable classification function that takes values from 0 to 9. For the training data, we had k = 58191 samples. Using a linear transformation, the 196-variable classification function was converted into a 25-variable function. We applied the testing data consisting of 9569 samples. The reduced classification function produced correct answers for 97.3% of the recognized test data. For unrecognized test data, the circuit for the reduced classification function produced "unrecognized" signals. The recognition circuit for handwritten digits can be implemented by a simple architecture: a cascade of a linear circuit and a memory. To increase the recognition rate, we also present methods using multiple classification functions and voters.