The History of AI

The neocognitron

Era 2 · AI is born (1943–1979) › Neural networks and learning

Lineage
Neurons and networks
Kind
Publication

Who

What happened

Kunihiko Fukushima publishes the neocognitron in Japanese in October 1979, in the IECE Transactions, and in English in April 1980, in Biological Cybernetics.

Why it mattered

A hierarchical network that alternates layers detecting local features with layers that tolerate shifts in their position, so that it recognizes a pattern even when it moves. It is the architecture behind convolutional networks.

The full story

In the late 1950s, David Hubel and Torsten Wiesel found that the cat's visual cortex contains simple cells, which respond to an edge with a certain orientation at a specific spot, and complex cells, which respond to the same edge at any nearby position. Fukushima, a Japanese engineer, turned that hierarchy into a network.

The neocognitron alternates two kinds of layer. S-cell layers apply the same feature detector at every position in the image, what is now called a convolution. C-cell layers combine the responses of neighboring positions and so tolerate small shifts, what is now called pooling. Stacking several such pairs makes the features progressively more complex and less dependent on position, until the network recognizes letters or digits even when they are shifted or somewhat distorted.

The network learned without supervision: the detectors organized themselves from the patterns presented, which Fukushima called learning without a teacher. It did not use backpropagation. Back in 1969, Fukushima had also described the rectified linear unit, the activation now known as ReLU.

The architecture of alternating convolutional and pooling layers is the one several groups trained with backpropagation in the late 1980s, Alex Waibel's on speech signals in 1987, Wei Zhang's in Japan in 1988, and above all Yann LeCun's in 1989, and the one that, scaled up, dominated computer vision from 2012 on.

A row of boxes from left to right, image, S1, C1, S2, C2, and category, each smaller than the last. The S layers are labeled "detects a feature at every position" and the C layers "responds even if the feature shifts."

S layers detect features and C layers tolerate their shifting; each pair combines broader features.

Common misconceptions

  • What people often say: Yann LeCun invented the convolutional network.

    In fact: The architecture of alternating detection and pooling layers is that of Fukushima's neocognitron, from 1980. What Waibel, Zhang, and LeCun did in the late 1980s was train it with backpropagation.

Videos

Builds on

  • Hubel, Wiesel, and the visual cortex (October 1959)

    Documented direct influence

    The S-cell and C-cell layers of the neocognitron take their names and their roles from Hubel and Wiesel's simple and complex cells

Leads to

  • LeCun's convolutional network (1989)

    Inherited, with no proven direct link

    The same hierarchical architecture; LeCun adds training by backpropagation

Threads that pass through here

Sources

How to cite

“The neocognitron.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/neocognitron

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline