Integrating CNN and RCE Networks for Efficient On-Device Image Classification

This paper introduces an algorithm that combines the feature extraction capabilities of Convolutional Neural Networks (CNN) with Restricted Coulomb Energy (RCE) networks to classify 2D images. Utilizing CNN as the feature extractor for the RCE network, the study evaluates performance using the MNIST dataset, demonstrating that the combined CNN + RCE network performs comparably to CNN alone. To test the on-chip learning capabilities of the RCE network, certain digits were excluded from the training process and later reintroduced during inference. This approach confirmed high classification accuracies of 92.11% and 96.23% for digits '0' and '1' respectively. The results validate the potential of the CNN + RCE network for on-device AI applications, with future work planned for hardware implementation and further performance evaluations.