This article delves into how gradient-based learning, particularly using multilayer neural networks and the back-propagation algorithm, can be effectively applied in the field of document recognition. This type of learning involves gradually adjusting the parameters of a neural network to improve its accuracy in tasks like recognizing handwritten characters.The key idea is that, with the right network architecture, these learning algorithms can create a complex decision-making process. This process is capable of classifying patterns that are high-dimensional and intricate, such as handwritten letters and numbers, with minimal initial processing of the data.The paper reviews various methods used for handwritten character recognition and compares them using a standard task of recognizing handwritten digits. Among these methods, convolutional neural networks (ConvNets) stand out. ConvNets are specially designed to handle the variability in 2D shapes, making them particularly effective for tasks like reading handwriting.Moreover, the paper introduces a new learning concept called Graph Transformer Networks (GTNs). GTNs are designed for more complex document recognition systems that involve several steps, such as extracting specific fields from a document, segmenting different parts of the text, recognizing the segmented parts, and understanding the language. GTNs enable all these different modules to be trained together in a way that optimizes the overall performance.The authors also discuss practical applications, including two systems developed for online handwriting recognition. These systems, trained globally using gradient-based methods, demonstrate the benefits of this comprehensive training approach. One notable application described is a system for reading bank cheques. It combines ConvNet-based character recognizers with these global training techniques, achieving high accuracy in reading both business and personal cheques. Impressively, this system is not just a theoretical model but is actually deployed in the real world, where it processes several million cheques every day.