Glossary
Who is Yann LeCun?
In short
Yann LeCun is an AI researcher who created convolutional neural networks and won the Turing Award for deep learning.
People · Yann LeCun
MMOOZZGG.com editorsChecked: 1 min read
For a child
How do you recognize a cat in a picture? First you notice small things: ears, whiskers, a tail. Then you put them together and understand who it is. LeCun taught a computer to look the same way. His program looks at small pieces of a picture, finds little lines and corners in them, and then builds whole letters and digits out of them. This is how a computer learned to read handwritten digits.
In plain words
LeCun found a way to teach a computer to see. An ordinary neural network looks at a picture as a long list of dots and gets confused if a digit moves a little. LeCun’s convolutional network passes a small window over the image and looks for the same patterns in every place: lines, arcs, corners. The next layers build parts out of them, and then whole characters. Such a network recognizes a digit even if it is written crooked or at a slant. His LeNet-5 system read handwritten and printed characters, and similar networks later became the basis of computer vision.
For business
Convolutional networks work wherever a company needs to read images. A warehouse scans delivery notes, a bank checks customer documents, a plant looks for defects in photos from the line, a store counts goods on shelves from photos. LeCun’s group described reading checks with such a network back in 1997, and document recognition is still one of the most common tasks. The cost of a project depends on the image quality, the number of document types and how many examples must be labeled for fine-tuning. The main risk: a model trained on clean scans does worse on crumpled phone photos. So before launch, teams collect real samples and count the error rate. The page on data for AI explains which data to start with.
For a specialist
A convolutional neural network (CNN) applies one shared set of trainable filters to an image, so the same feature is recognized anywhere in the frame, and it needs fewer parameters than a fully connected network. Pooling layers shrink the feature map and give stability against shifts and distortions; training uses backpropagation. The LeNet-5 architecture from the paper by LeCun, Bottou, Bengio and Haffner, “Gradient-based learning applied to document recognition” (Proceedings of the IEEE, November 1998), recognized handwritten and printed characters. In 1997 the same group described reading checks with graph transformer networks. LeCun shared the 2018 Turing Award, announced on 27 March 2019, with Geoffrey Hinton and Yoshua Bengio for breakthroughs in deep learning and convolutional networks. At the time of the award he was a professor at New York University and vice president and chief AI scientist at Facebook.
Example
The accounting team gets hundreds of invoice scans a week. A convolutional network finds the fields with the amount, the date and the tax ID on each scan, and an accountant checks only the documents the model is unsure about.
Related terms
Questions and answers
Sources
Not sure whether your company needs this?
AI solutions map