LeNet

LeNet is one of the early neural networks that showed computers could learn to read images rather than relying entirely on hand-written rules. It became famous for recognizing handwritten digits—the kind found on bank checks and postal codes—and helped establish the design ideas behind modern image-recognition systems.

How LeNet reads an image
The best-known version, LeNet-5, takes a small grayscale image, such as a 32×32 pixel digit, and processes it through a sequence of learned layers. Its key idea is the convolutional layer: small learned filters slide across the image looking for useful visual patterns. Early filters respond to simple marks such as edges, curves, and corners; later layers combine these into digit-like parts and complete shapes.

The main building blocks

• Convolution layers create feature maps that preserve where visual patterns occur.
• Pooling layers shrink those maps, reducing computation and making the network less sensitive to a feature shifting by a few pixels.
• Fully connected layers use the extracted features to choose a class, such as “3” rather than “8.”

Why it still matters
LeNet is small by current standards, but its pattern—convolutions, spatial downsampling, then classification—became the foundation for later CNNs such as AlexNet and VGG. In practice, a closely related model can classify simple grayscale images, inspect characters in scanned documents, or recognize marks in controlled manufacturing images. Modern libraries make its structure easy to build; for example, PyTorch provides Conv2d, AvgPool2d, and Linear layers for the same core pipeline. Without learned convolutional features, a system would need manually designed rules for every edge, stroke, and variation in appearance—rules that break quickly on messy real images.