Research · November 1998
LeNet-5 and MNIST
The paper set out a complete convolutional architecture for document recognition and, with it, a handwritten digit set that became a shared ruler.
Why it matters
The architecture and the dataset AlexNet was built on fourteen years later appear here in finished form.
LeNet-5 has seven layers of convolution and subsampling and trains end to end by gradient, from pixels to answer. The authors assembled the 70,000-image MNIST set from United States census databases. It is still the most common teaching example. The paper also argues a broader thesis: a document is better recognised by one trained system than by a chain of separately tuned steps.
Event record
- Event date
- November 1998
- Timeline date
- Event date
- Verification
- Sources gathered automatically · September 17, 2026
- Lines
- ID
- evt-0178
Proceedings of the IEEE volume 86, issue 11, November 1998.
Records that link to this one
- Enables NIST's shared test on handwritten characters
MNIST was built from NIST Special Database 3, the training disc of this conference, and Special Database 1.
THE MNIST DATABASE of handwritten digits (Yann LeCun, Corinna Cortes)The First Census Optical Character Recognition System Conference (R. Allen Wilkinson, Jon Geist, Stanley Janet, Patrick J. Grother, Christopher J. C. Burges, Robert Creecy, Bob Hammond, Jonathan J. Hull, Norman J. Larsen, Thomas P. Vogl, Charles L. Wilson), NISTIR 4912, August 1992 - Related A convolutional network reads cheques in banks
The network reading the cheques is LeNet-5, whose full account appeared the next year.