This paper presents a comprehensive review of gradient-based learning applied to document recognition, focusing on the use of multilayer neural networks and convolutional neural networks (CNNs) for handwritten character recognition and document understanding. The authors argue that better pattern recognition systems can be built by relying more on automatic learning and less on hand-designed heuristics. They demonstrate that CNNs, which are specifically designed to handle the variability of two-dimensional shapes, outperform other techniques in handwritten digit recognition. The paper also introduces graph transformer networks (GTNs), which allow multimodule systems to be trained globally using gradient-based methods to minimize an overall performance measure. Two systems for online handwriting recognition are described, showing the advantage of global training and the flexibility of GTNs. A GTN for reading bank checks is also described, which uses CNN character recognizers combined with global training techniques to provide record accuracy on business and personal checks. The paper discusses various learning techniques, including gradient-based learning, back-propagation, and the use of convolutional networks for isolated character recognition. It also explores the use of GTNs for recognizing handwritten and machine-printed bank checks, with the core of the system being the convolutional NN called LeNet-5. The paper concludes that gradient-based learning, particularly with CNNs and GTNs, is a powerful approach for document recognition and pattern recognition tasks.This paper presents a comprehensive review of gradient-based learning applied to document recognition, focusing on the use of multilayer neural networks and convolutional neural networks (CNNs) for handwritten character recognition and document understanding. The authors argue that better pattern recognition systems can be built by relying more on automatic learning and less on hand-designed heuristics. They demonstrate that CNNs, which are specifically designed to handle the variability of two-dimensional shapes, outperform other techniques in handwritten digit recognition. The paper also introduces graph transformer networks (GTNs), which allow multimodule systems to be trained globally using gradient-based methods to minimize an overall performance measure. Two systems for online handwriting recognition are described, showing the advantage of global training and the flexibility of GTNs. A GTN for reading bank checks is also described, which uses CNN character recognizers combined with global training techniques to provide record accuracy on business and personal checks. The paper discusses various learning techniques, including gradient-based learning, back-propagation, and the use of convolutional networks for isolated character recognition. It also explores the use of GTNs for recognizing handwritten and machine-printed bank checks, with the core of the system being the convolutional NN called LeNet-5. The paper concludes that gradient-based learning, particularly with CNNs and GTNs, is a powerful approach for document recognition and pattern recognition tasks.