In 1998, Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner published “Gradient-Based Learning Applied to Document Recognition” in the Proceedings of the IEEE (volume 86, issue 11, pages 2278 to 2324). The publisher record confirms the title, the four authors, the venue, and the year.
The paper presented LeNet-5, a convolutional neural network trained end to end to recognize handwritten characters, and argued that a system should learn its own features from raw pixels rather than relying on hand-engineered rules. It is one of the most thorough early demonstrations that a single network, trained by backpropagation, could read real documents such as bank checks at scale.
LeNet-5 is the direct ancestor of modern image recognition. Its core ideas, convolution, pooling, and gradient training, are the same building blocks that powered AlexNet in 2012 and the deep-learning vision boom that followed. The 1998 paper proved the architecture worked years before the data and computing power existed to make it dominate.