LeNet

LeNet is a family of convolutional neural networks developed by Yann LeCun and collaborators at Bell Labs for handwritten digit recognition. The most influential variant, LeNet-5, was introduced in the late 1990s and demonstrated end-to-end learning from raw pixel data to digit labels. LeNet helped establish convolutional neural networks as a practical tool for pattern recognition and remains foundational in the history of deep learning.

LeNet-5 processes greyscale images, typically 32 by 32 pixels. Its architecture combines convolutional layers, subsampling (pooling) layers, and fully connected layers. The common configuration begins with C1: six 5×5 convolutional kernels producing six 28×28 feature maps, followed by S2: a 2×2 average pooling operation reducing size to 14×14. Next, C3 applies 16 5×5 convolutions to yield 10×10 maps, then S4 performs 2×2 pooling to 5×5. C5 applies 120 5×5 kernels to the 5×5 input, producing 1×1 feature maps. The network then uses a fully connected layer (F6) with 84 units and a final 10-unit output layer for digit classes 0–9. The activation is typically tanh, and training is performed with backpropagation.

LeNet-5 incorporated local receptive fields and weight sharing, and used a relatively small parameter count by modern standards. It was trained on MNIST and demonstrated robust hand-written digit recognition. The work, published in the late 1990s, is frequently cited as a milestone in the development and popularization of convolutional neural networks, influencing later architectures and the broader adoption of deep learning.