PAPERS-02-002 · Paper — The First Convolutional Network (LeNet)

The paper

The First Convolutional Network (LeNet) by LeCun, Bottou, Bengio, and Haffner (1998, Proceedings of the IEEE) is a landmark because it introduced a practical, trainable CNN that could recognize handwritten digits end‑to‑end, setting the template for modern deep vision systems.

Before this paper

Before LeNet, visual pattern recognizers relied on hand‑crafted features and shallow classifiers that struggled with translation variance and required extensive feature engineering. Large‑scale digit recognition was handled by template matching or statistical models that could not learn hierarchical representations. Computing resources and labeled data were insufficient to train deep models, so the field lacked a scalable, learnable architecture for images.

The key idea

LeNet popularized three core ideas: local receptive fields that restrict each neuron to a small patch of the image, shared weights that let a single filter slide across the whole input, and subsampling (pooling) that provides translation tolerance. These concepts together formed a hierarchical feature extractor that could be trained by gradient descent. The network demonstrated that convolution and weight sharing could dramatically reduce the number of parameters while preserving expressive power.

How it works

An input image first passes through a convolutional layer where a set of learned filters scan the image, producing feature maps that capture local patterns. Each feature map is then down‑sampled by a subsampling (pooling) layer, reducing spatial resolution and providing robustness to small shifts. The pooled maps feed into additional convolution‑pooling stages, building increasingly abstract representations. Finally, fully‑connected layers map the high‑level features to digit class scores, and the whole system is trained end‑to‑end with backpropagation using labeled MNIST examples.

See it — the LeNet-5 pipeline

A schematic of LeNet‑5 showing alternating convolution, subsampling, and fully‑connected layers.

flowchart LR
  A["32x32 digit"] --> B["conv + subsample<br/>learned filters"]
  B --> C["conv + subsample<br/>deeper features"]
  C --> D["dense layers"]
  D --> E["10 digit classes"]

What they showed

On the MNIST handwritten‑digit benchmark, LeNet‑5 achieved about 0.8% test error, demonstrating that a shallow CNN could reach near‑human performance on a real‑world recognition task. The model was also deployed commercially to read handwritten digits on bank checks in the late 1990s, processing a substantial share of US checks at the time.

What it changed

LeNet proved that convolution, weight sharing, and pooling could be learned jointly, inspiring a generation of deeper CNNs that dominate computer vision today. Its success showed that end‑to‑end training was feasible, paving the way for later breakthroughs once data and compute caught up. The architecture became the de‑facto reference for subsequent models, influencing frameworks, research curricula, and industrial deployments.

Next: The Moment Deep Learning Won (AlexNet).

The one-line takeaway

LeNet turned the vague idea of visual filters into a concrete, trainable engine that still powers modern vision systems.

Read it / meet it

Read the original 1998 IEEE paper for the full algorithmic details and training setup. Today you can see LeNet‑5 implemented in every deep‑learning library (TensorFlow, PyTorch, Keras) as a classic example and tutorial model.