Arabic handwritten digit classification without gradients: Pseudoinverse Learners

Abstract

The topic of handwritten digit recognition (HDR) has drawn more and more attention in recent years. The biggest drawback of HDR is the lack of an efficient model that can categorise the handwritten numbers that users present via digital devices. Various methods have been developed to enhance HDR in Arabic, using on sophisticated deep learning techniques such convolution neural networks (CNNs), which are learning using a backpropagation algorithm that has many drawbacks, including: 1) Local minima possibility, 2) long learning time, 3) Non-assured convergence, 4) Selective learning data and 5) Black box: the inner mapping techniques of the BP are remaining unclear and not grasped. To overcome these limitations, this paper introduces the use of pseudoinverse learning autoencoder (PILAE) algorithm. The PILAE is not a gradient descent strategy; however, it is not required to specify the learning rate or suggest the quantity of hidden layers and the drawback of gradient vanishing. According to experimental findings, the introduced technique improves test accuracy while maximising computational efficiency.