The History of AI

GMDH

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

Lineage
Neurons and networks
Kind
Implementation

Who

Institute of Cybernetics, Kyiv

What happened

In Kyiv, Alexey Ivakhnenko and Valentin Lapa publish a method for training multilayer networks by adding layers one at a time, the Group Method of Data Handling (GMDH).

Why it mattered

According to Schmidhuber's history of deep learning, it is the first general, working algorithm for training networks with many layers. In 1971 Ivakhnenko describes an eight-layer network.

The full story

In the Soviet Union of the 1960s, cybernetics was a prestigious field, and in Kyiv the engineer Alexey Ivakhnenko worked on models for predicting and controlling complex processes from data. The approach was practical: rather than assume a form for the model, let the data select it.

The method builds the network layer by layer. Each unit combines a few inputs with a low-degree polynomial whose coefficients are fitted by regression. The units are then evaluated on data not used to fit them, the best ones are kept, and a new layer is added on top. Growth stops when the network no longer improves on that validation data, an early way of avoiding overfitting.

The result is a deep network in the literal sense, with layers that learn intermediate representations. A 1971 paper by Ivakhnenko describes one with eight layers. The method was applied to prediction, optimization, and pattern recognition.

It remained almost unknown in the West. In a history of deep learning, Schmidhuber presents it as the first general, working algorithm for many-layered perceptrons, two decades before the term deep learning caught on.

Sources

How to cite

“GMDH.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/ivakhnenko-gmdh

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline