Kunihiko Fukushima proposed the Neocognitron in 1980 as an early precursor to convolutional neural networks.
The architecture used alternating layers with different roles: some detected visual features, while others helped make recognition less sensitive to where those features appeared in an image. This hierarchical design was inspired by ideas about visual processing in the brain.
Neocognitron was designed for visual pattern recognition, including recognizing handwritten or distorted characters. Its use of local feature detectors and progressively more abstract representations anticipated important ideas later used in modern computer vision.
It is often described as a predecessor of LeNet-5, the convolutional network developed by Yann LeCun and colleagues in the 1990s. However, Neocognitron was not identical to later CNNs: its learning method and layer design differed from the backpropagation-based systems that became widespread. The name can also be confused with the perceptron, which is an earlier and much simpler neural-network model rather than a hierarchical visual architecture.