Yann LeCun, who at AT&T Bell Labs in 1989 first trained convolutional networks end to end with backpropagation (later LeNet). The architecture builds on Kunihiko Fukushima's 1980 neocognitron.
Card 1 of 259. Answer: Yann LeCun, who at AT&T Bell Labs in 1989 first trained convolutional networks end to end with backpropagation (later LeNet). The architecture builds on Kunihiko Fukushima's 1980 neocognitron.
About · Quizzes · History of AI · To suggest new questions or report an error send me a dm.
Yann LeCun is usually called the founding father of convolutional networks (CNNs). In 1989, at AT&T Bell Labs, he and colleagues trained a convolutional network end to end with backpropagation and put it to real work reading handwritten zip codes. A CNN applies small learned filters, shared across positions, over an image.
The architecture is older. Kunihiko Fukushima's neocognitron, published in Biological Cybernetics in 1980 and modeled on Hubel and Wiesel's simple and complex cells, already stacked feature-detecting S-cell layers and shift-tolerant C-cell layers, the ancestors of convolution and pooling, but it learned by unsupervised self-organization, not backpropagation. LeCun's 1989 network read 16 × 16 digit images: thanks to weight sharing, its 64,660 connections needed only 9,760 free parameters, and it reached 5% test error after training on 7,291 digits.
So the precise answer credits Fukushima for the design and LeCun for making it trainable and practical, leading to LeNet-5 in 1998. Interviewers often ask what weight sharing buys: far fewer parameters, and the same detector applied at every position.
