> LeNet_

In 1989, Yann LeCun, a young French researcher working at Bell Labs in Holmdel, New Jersey, published a paper that demonstrated convolutional neural networks (CNNs) could recognize handwritten digits with remarkable accuracy. The system, which came to be known as LeNet, was trained on images of handwritten ZIP codes collected from envelopes processed by the U.S. Postal Service. It could read digits with an error rate below 1% — better than any previous system and approaching human performance. For the first time, a neural network had proven itself on a practical, real-world task that mattered to a large organization with a very concrete bottom line.

The architecture of LeNet embodied a beautiful insight about how vision works. Instead of treating an image as a flat array of pixels, LeNet used "convolutional" layers that applied small filters across the image, detecting local features like edges, curves, and corners. These features were then combined in successive layers to recognize increasingly complex patterns — edges into strokes, strokes into shapes, shapes into digits. The network also used "pooling" layers that provided a degree of translation invariance, meaning it could recognize a "7" whether it appeared in the upper left or lower right of the image. The entire architecture was inspired by the work of Hubel and Wiesel, who had won the Nobel Prize in 1981 for discovering that the mammalian visual cortex processes images through a hierarchy of increasingly complex feature detectors.

LeNet was deployed at scale. AT&T and NCR used LeCun's system to process handwritten checks at ATMs and bank clearing houses. By the mid-1990s, LeNet and its successors were reading approximately 10% to 20% of all the checks deposited in the United States — millions of checks per day. This was neural network technology operating in the real world, handling real money, and doing so reliably enough that banks trusted it with their core operations. It was not a research demo or a proof of concept; it was a production system that processed billions of dollars in transactions.

Yet LeNet's success did not spark an immediate revolution. The second AI winter was in full force, and neural networks remained deeply unfashionable in mainstream AI research. Funding agencies and top journals were skeptical. LeCun himself recalled that throughout the 1990s and early 2000s, getting papers on neural networks accepted at major machine learning conferences was extraordinarily difficult. The dominant paradigm was support vector machines (SVMs) and other kernel methods, which had elegant mathematical properties and worked well on small datasets. LeCun, along with Geoffrey Hinton and Yoshua Bengio, kept the faith through a long wilderness period. They continued refining their networks, publishing papers, and training students, even as the broader community dismissed them. When the deep learning revolution finally arrived in 2012, LeNet's architecture was its direct ancestor — the convolutional neural network that Alex Krizhevsky used to win the ImageNet competition was essentially LeNet on steroids, scaled up with GPUs and trained on millions of images.