Kunihiko Fukushima is a Japanese computer scientist whose neocognitron, published in 1980, is widely regarded as the direct conceptual ancestor of the modern Convolutional neural network (CNN), the architecture that underlies most of today's Computer vision systems.
The neocognitron
Born in 1936, Fukushima spent much of his career at NHK's Science and Technical Research Laboratories in Japan before later academic appointments, including at Osaka University. Drawing on the neurophysiological findings of David Hubel and Torsten Wiesel about simple and complex cells in the visual cortex of cats, Fukushima designed a multilayered Neural network in which alternating layers of feature-detecting "S-cells" and position-tolerant "C-cells" allowed the network to recognize handwritten characters and simple shapes regardless of small shifts or distortions in the input image. This hierarchical structure, in which each layer built more abstract features from the layer below and pooling operations provided a degree of translation invariance, anticipated the convolution-and-pooling design later used in CNNs.
The neocognitron was trained using an unsupervised, layer-by-layer learning rule rather than end-to-end Backpropagation, which had not yet been popularized for multilayer networks. It could not be trained as efficiently or scaled as far as later architectures, but its structural ideas proved durable.
Influence
Yann LeCun has repeatedly credited the neocognitron as a direct influence on LeNet and the broader CNN family he developed at Bell Labs in the late 1980s and 1990s, particularly the idea of local receptive fields and pooling. That lineage runs through the CNNs used in the 2012 AlexNet result, which triggered the mainstream turn to Deep learning, and continues in the architectures used across modern computer vision and parts of Multimodal AI systems today. Fukushima has received recognition including the IEEE Neural Networks Pioneer Award and the 2021 Bower Award and Prize for Achievement in Science, and remains one of the field's clearest examples of biologically inspired computing directly shaping a dominant engineering architecture, long before the limitations identified in early Perceptron research were fully resolved by later gradient-based training methods. Unlike many of the researchers who later built on his ideas, Fukushima worked largely within Japanese industrial and academic labs rather than the American university and corporate-lab networks that produced most of the field's subsequent celebrated breakthroughs, a fact often noted when his relatively limited public recognition outside specialist circles is discussed.