Yann LeCun, with Bengio, Bottou, and others, showed that convolutional nets trained with backprop could read handwritten digits at industrial scale—including bank-check systems.
CNNs baked in useful invariances (local translation) via shared filters, cutting parameters versus dense MLPs. The MNIST dataset became a standard testbed.
For years CNNs were “too expensive” for large images. ImageNet + GPUs changed that in 2012, but the seed architecture was already in place in the 1990s.
