A Comprehensive Timeline of Artificial Intelligence from the 17th Century to the Present

Jürgen Schmidhuber recently compiled a detailed AI history that spans from the 17th century to modern deep learning, providing a timeline of pivotal events, inventions, and influential scientists.

The term "Artificial Intelligence" was coined at the 1956 Dartmouth conference, but practical AI concepts date back to 1914 when Leonardo Torres y Quevedo built the first working chess‑playing machine.

Early theoretical foundations were laid in the 1930s by Kurt Gödel, who identified fundamental limits of computation‑based AI.

In the 1980s, research emphasized theorem proving, logic programming, expert systems, and heuristic search.

The 2000s saw a focus on support‑vector machines, Bayesian reasoning, decision trees, ensemble methods, swarm intelligence, and evolutionary computation, which powered many successful AI applications.

During the 2020s, AI research returned to fundamentals such as the chain rule and gradient‑descent training of deep nonlinear neural networks, especially recurrent networks.

Key historical milestones include:

1676 – Gottfried Wilhelm Leibniz published the chain rule, later becoming the credit‑allocation core of modern deep learning.

1805 – Adrien‑Marie Legendre described what is now recognized as a linear neural network (least‑squares regression).

1920‑1925 – Ernst Ising and Wilhelm Lenz introduced the first non‑learning recurrent neural network (the Ising model), later extended to a learnable RNN by Shun‑Ichi Amari in 1972.

1979 – Kunihiko Fukushima created the Neocognitron, the prototype of today’s convolutional neural networks (CNNs).

1990 – The earliest notion of generative adversarial networks (GANs) appeared under the name "AI curiosity".

1991 – Seppo Linnainmaa published the back‑propagation algorithm, which underlies modern deep learning frameworks.

1995 – Schmidhuber introduced a neural probabilistic language model.

1997 – Long Short‑Term Memory (LSTM) networks were introduced, later becoming the most cited neural network paper.

2005 – Connectionist Temporal Classification (CTC) enabled end‑to‑end speech recognition with LSTMs.

2015 – Highway Networks and ResNet (derived from Highway Nets) demonstrated that very deep feed‑forward networks could be trained effectively.

2018 – A policy‑gradient‑trained LSTM powered OpenAI’s Dactyl robot hand.

2019 – DeepMind’s AlphaStar, built on a deep LSTM core, defeated professional StarCraft players.

Schmidhuber also emphasized the importance of self‑supervised neural history compressors and sub‑goal generators, which allow hierarchical abstraction and planning across multiple time scales.

Hardware advances have been crucial: from Leibniz’s mechanical calculators to modern GPUs containing billions of transistors, enabling the massive parallelism required for deep learning.

Physical limits such as the Bremermann limit suggest that future efficient computing will need brain‑like three‑dimensional, densely packed, sparsely connected RNN architectures.

Overall, the article argues that deep learning’s core is network depth, with LSTM and Highway/ResNet architectures forming the backbone of contemporary AI systems across vision, language, speech, and reinforcement learning.

Reference: https://people.idsia.ch/~juergen/deep-learning-history.html