RFF based Enhanced CNN Architectures for Handwritten Alphabet/Digit Recognition Using the EMNIST Dataset

Abstract

Reproducing kernel Hilbert Space based machine-learning/deep learning methods have emerged as a promising solution for learning in low-data regime, and for enhancing shallow networks for generic datasets. This paper shows the application of random Fourier features (RFF) based neural networks to the extended MNIST (EMNIST) dataset for handwritten character/digit recognition. From our simulations, we find that RFF-enhanced shallow convolutional neural networks (CNN) deliver improved performance compared to shallow CNNs in the low-data regime. Furthermore, with shallow RFF-CNNs, we observe similar performance metric compared to very deep CNNs such as ResNet and AlexNet, which lowers the overall computational overhead. Furthermore, by comparing a pruned AlexNet with its RFF-CNN enhanced counterpart, we show the efficacy of RFF-CNN in transfer learning. Lastly, using small number of RFF, a computationally simple filtered RFF-CNN (F-RFF-CNN) architecture is proposed which generalizes equivalently to an RFF-CNN with a large number of RFF. The above RFF-CNN architectures are validated by simulations performed using the 62 class (a-z,A-Z,0-9) EMNIST dataset, which makes the paradigms of RFF-CNN and F-RFF-CNN promising for character/digit recognition. © 2026 IEEE.