As predicted in 1984, a second period of disinterest in AI and reduced funding occurred in 1987. Its causes were a mix of the hype cycle turning downwards and market dynamics :

• Expert Systems (the epitomy of top-down AI) showed serious limitations. Being optimized for specific knowledge and context, they became useless out of those parameters.  The rise of neural network shifted attention elsewhere. In 1987, Alactrious Inc launched Alacrity, an Expert system using over 3000 rules, which was one of the last of its kind.
  
• Expensive, complex, computers build spefically to run expert systems were mae redundant by cheaper, more powerful, generic workstations, that could also run LISP and expert systems (SUN, IBM, Apple). A whole industry collapsed in the year of 1987.
  
• Japan's Fifth Generation program came to an end, to be replaced by a 6th generation dedicated to neural-nets. Results didn't come close to initial hopes, and all had to be rebuilt in a different direction, which would take years.
  
• DARPA, which had started to fund AI again in 1983, to follow Japan's lead, decided once again that AI did not deliver on its promises and directed funding to alternative fields of research, which did.  
  

1986: VaMoRs, Enrst Dickmanns' robotic vehicle based on a Mercedes van, becomes the first autonomous vehicle able to drive itslef. By the next year, it would drive up to 60mph on empty roads. And by 1994, it was tested (in Mercedes 500 form) on public French roads at speeds up to 130km/h.

1986: Geoffrey Hinton and his colleagues publish an article about Backpropagation. This algorithm is used to adjust the weights of connections between individual neurons in neural networks to minimize the difference between the network's prediction and the exact answer. 

Before Backpropagation, the potential of neural networks was understood, but training them was messy and complex business. Backpropagation introduced a systematic approach that made the training much more efficient and easy. Until then, top-down AI approaches such as experts systems had been more prominent, but this sparked renewed interest in bottom-up approaches such as neural networks, which learn not from rules but from examples.

1986: UCI Professor Rina Rechter coins the term "Deep Learning".

Not all research stopped during the first AI winter, and some good news began to thaw the ice towards the end of the decade. 

1979: The "American Association for Artificial Intelligence" is founded, signaling a structuration of the academic community. It was a nonprofit scientific society devoted to "advancing the scientific understanding of the mechanisms underlying thought and intelligent behavior and their embodiment in machines", renamed to the current "Association for the Advancement of Artificial Intelligence" (AAAI) in 2007.

1979: Kunihiko Fukushima (father of ReLU) creates a form of multilayer neural networks callad Convoluted Neural Networks (CNN) still in use today in image recognition. The subsequent resurgence of interest in AI in Japan, particularly the Fifth Generation government program, got other governments back in the race, helping end AI Winter n°1.

1980: DEC releases R1 - known internally as XCON, for eXpert CONfigurer - the first ever commercial expert system, used to guider potential buyers of DAC's VAX computer solutions. This commercial dimension also contributed to end the first AI winter.

The enthusiastic scientists and media of the time forgot a crucial rule in business (in this case, funding): underpromise and overdeliver. Officials in the various governments and agencies funding AI research began to doubt that advances matched espectations, among other problems: inoperational machine translation, limitations of perceptrons, speed issues. DARPA started cutting finances of projects that had no direct application, feeling the no-strings-attached funding of theoretical research was getting them nowhere. 

The UK government, via the Science Research Council, commissioned British Mathematician and innovator James Lighthill to asses the field's progress.

Lighthill's report was scathing. Having conducted a compilation of the academic research in AI and robotics, he concluded that most research had ignored complexity theory ("combinatorial explosion") and was unfit to address real-life situation. Consequently, funding was stopped for all but 2 British universities.

And the domino effect that caused other funding agencies to follow suit was compared to the chain reaction that occurs in nuclear explosions, causing nuclear winters in the atmosphere. So the field enterted its first major pause after decades of optimism. The first AI winter. At this time, public interest also waned.

WAseda roBOT 1, a.k.a. WABOT 1, had legs. It was the first robot with slightly antropomorphic appearance. This was deemed important if robots were to enter society. It was designed to function in a human environment and interact with human beings. 

It was built by Waseda University in Japan around 1972, and could communicate in Japanese, walk, measure the environment around it to avoid obstacles, and could carry objects with artificial hands. 

While Marvin Minsky (see quote below) enthusiastically - and somewhat optimistically - promised robots with human intelligence within the 1970s, WABOT 1 was estimated to possess the IQ of a 1 year-old child.

Alexey Ivakhnenko was a Ukranian computer scientist who specialized in pattern recognition, forecasting in complex systems and self-organisation. He developed a technique called inductive learning which was a direct competitor to the more mainstream stochastic approaches.

In 1968, he published an important article in the journal "Avtomatika" about Group Method of Data Handling (GMDH), a method which competed with the traditional deductive processes of science. Instead of working (top-down) from a theory to a model of it, GMDH proposed to work (bottom-up) from the observation of data to a model which might explain the data.

This work not only proposed to change the way general science could be carried out, but also offered a new direction to Artificial Intelligence, away from Expert Systems and towards Machine Learning as we know it today.

He is considered the father of Deep Learning.  

Remember McCulloch and Pitts further down the page, from 1943? Their article "A Logical Calculus of Ideas Immanent in Nervous Activity” theorised the artificial neuron and provided a mathematical foundation for connectionism (the neural netwoks approach to AI)

In 1957, American psychologist created a program to emulate it on the IBM 704 computer, then built a hardware implementation in 1958, designed for image recognition with a 400 photosite array (20 x 20 Pixels, yoohoo).

In Machine Learning terms, a perceptron is a supervised learning binary classifier. It takes a number of inputs (akin to the dendrites in human neurons) and performs a calculation that answers Yes or No.

Media reception in 1958 was extremely enthusiastic, suggesting machines would soon "walk, talk, see, write, reproduce itself and be conscious of its existence", no less. But experimentation soon revealed the limitations of the single perceptron, leading to the understanding of the need for more elaborate neural networks.

1956 is a very important year for the growth of Artificial Intelligence. 

First, Allen Newell, Herbert A. Simon, and Cliff Shaw write "Logic Theorist", the very first program aimed at automated reasoning. The program was able to prove 38 of the 52 mathematical theorems in a section of Principia Mathematica, even finding more elegants proofs for some of them. It is regarded as the first true Artificial Intelligence Program. Among other important concepts, it used search trees and heuristics.

And, over the summer of 56, researcher John McCarthy organised an 8 week workshop at Dartmouth university, which is now consered the first academic event for this field. Attendees included most of the famous names of the time in Computer Science and the term Artificial Intelligence was coined during the proposal of this event.

Artificial Intelligence was officially born and, with it a period of unbridled optimism that would last almost two decades.

Gamers of the world, unite and bow to your master 😉 Employed at IBM, Arthur Samuel had an obsession with developing a program that would exhibit actual intelligence. He programmed the IBM 701 computer to play draughts (checkers), which is a simple to program game but still benefits from strategy and thought. 

The significance of this program stems from the learning strategy Samuel implemented. Every game played with a human being contributed to improved performance (similar to today's Reinforcement Learning strategies). Over time, the program got better and eventually beat Samuel! For this, Samuel is considered the true father of Machine Learning. 

Alan Turing worked on cognitive science and computer science. His research helped launch connexionism (neural networks) and "artificial life". One of his major contributions came in the form of the Turing test, in which an observer of a conversation between a machine and a human being would have to determine which is the machine. If humans are unable to detect the machine, it is said to pass the test.

Simultaneously, Artificial Intelligence was powering the imagination of science-fiction authors such as Isaac Asimov. His book I, Robot introduced the now famous 3 laws of robotics intended to protect mankind from intelligent robots. The topic is still very fresh over 70 years later.

Interestingly, some research may have been influenced by the imagination of science fiction authors, though this it not widely documented, and may only be a pleasant idea 🙂

It was designed in 1948, by Frederic C. Williams and Tom Kilburn, at Manchester University (UK), and ran its first complex program in 1949. It's purpose was to elevate computer literacy in the university, but also to serve as a prototype for the commercial development of the Ferranti Mark I, the first commercial computer in history, released in 1951. 

Its main differenciator with respect to the previous generation was the presence of memory registers which could store the instructions of an algorithm (rather than have them on punched cards or via dedicated cabling, for instance). Several similar designs followed closely at Cambridge University (UK) and in the US army.

In 1943,  American neurophysiologist and cybernetician Warren Sturgis McCulloch, and American logician Walter Pitts, published a seminal paper that has received little attention from the AI community but was the first to use logic and computer science approaches to mental (neural) activity.  

In that respect, it layed the ground work for the first artificial neuron, the Perceptron which would be announced in 1957. And, more importantly, it attempted to prove that a network finite number of such (formal, not biological) neurons could implement a Turing machine.

As such, "A Logical Calculus of Ideas Immanent in Nervous Activity” was probably the first scientific article on Artificial Intelligence.

From the Pascaline to Alan Turing's message decyphering machines of WWII (the "bombes") all calculators had been (1) electromechanical and (2) designed to perform exactly one task, with human input. The next generation of computing machines would change that.

In 1941, Konrad Zuse designed the Z3 electromechanical computer, that was both programmable (so able to perform different tasks) and digital (its relays immplemented Boolean logic). In spite of a limited instruction set, it was proved to be Turing-complete, ie able to implement a Turing Machine, and run any algorithm. Zuse's requests for funding to replace electromechanic relays with electronic switches was turned down by the German government during WWII.

In 1945, IBM built the Automatic Sequence Controlled Calculator (ASCC), a.k.a. Harvard Mark I, designed by Howard Aiken, who took some inspiration from Charles Babbage's Analytical Engine (which could never be implemented before). It was still based on electromechanical technology but was fully programmable and required no human intervention after programming. One of the first users was John Von Neumann, during the Manhattan Project.  

And, in 1946, ENIAC (Electronic Numerical Integrator and Computer) became the first machine to use electronic tubes instead of electromechanical relays. It was designed to calculate ballistic trajectories for the US army, a task it could complete in 3 seconds instead of over 2 days using Babbage's machine and 2 hours with Harvard Mark 1. But the first user was again John Von Neuman whom it helped design the hydrogen bomb. It was fully programmable, but needed recabling for every new program, and tubes heated up and were notorious for their lack of reliability.

In spite of the increases in speed and autonomy this trio of machines brought to the world of computing, they still suffered from three issues severely hindering their use for Artificial Intelligence. They were still too slow, too expensive and couldn't store the programming instructions internally, in memory. This would soon change again.

Alan Turing's contributions were both theoretical and very down to Earth. After Goedel (in 1931) published a theorem proving that the then in-vogue dream of a "theory of everything" was impossible (the Incompleteness Theorem), Alan Turing (and his PhD tutor Alonzo Church) proved that some problems are not even "decidable", ie you can't even decide whether a solution can be found or not. In 1935, he designed a simple abstract machine (the TuringMachine) capable of implementing any algorithm a modern computer can.

His crucial work at Bletchley Park to break the German Enigma code is now legend. The interesting part, in the context of this page, is the electromechanical machines he designed and helped build to speed up the number crunching required for this task. After the war, he went on to design the first electronic stored-program general-purpose digital computer.

In 1950, he designed the famous Imitation Game, a test of machine's abilities to think in a way that is undistinguishable from human beings. Recent deep learning models such as GPT are now challenging this test. 

Just as mechanical devices had been invented because manual calculations were too slow, Hollerith introduced electricity to punched card calculators to speed up calculations again. 

As a statistician, his need was to tabulate large amounts of data for the United States Census Bureau.

His improvement on current systems was made through the use of electic connections to increment counters, rather than manual handling. And based on Boolean concepts of True or False, he used holes in punched cards to indicate specific characteristics of a person in a census (hole: married, no hole: unmarried, for example). Each card used the exact same layout to describe the various information about individual people.

Using these punched cards in his electric calculator made the calculation of population statistics much faster than would have been possible previously.

In 1896, the entrepreneurial Hollerith founds the Tabulating Machine Co, which was headed in
 1914 by Thomas J Waston, and was renamed International Business Machines Corporation
in 1917 IBM was born!

If Jacquard, below, partly paved the way for the industrial revolution, English polymath George Boole played as significant a role in laying the foundations for the digital revolution and Computer Science (of which AI is a branch).

His major contribution in mathematics is boolean algebra, in which a variable is either true or false, and operations combine those values in clearly defined ways. His work was instrumental in separating logic from philosophy (the two fields were strongly connected at the time) and helped define symbolic logic, in an attempt to formalise the brain's reasoning.

He was also entirely self-taught, something equally possible today in the field of Artificial Intelligence. 

English engineer Charles Babbage described the first principles of what modern computers would become, hoping to create more accurate nautical, astronomical and mathematical tables than humans of the time could achieve. The design was close to the combination of mechanical calculators and punched cards. Babbage built a model of his computer, but was never able to complete it owing to high costs and falling support.

Young Ada Lovelace was fascinated by early designs and it was her idea to add the punched cards, so as to be able to program the machine rather than simply use it manually. She is credited with the first ever algorithm (evidence suggests Babbage created some too), the set of instructions to be punched into the cards to perform complex mathematical calculations. 

Just as in AI today, she envisioned uses of computers that weren't limited to maths, but extended to the creation of music and poetic mathematics. Read about her extraordinary life and contribution to Babbage's machine.

By using a sequence of punched cards (invented decades before by Mrs Bouchon, Falcon and Vaucanson) Joseph Marie Jacquard was able to bolt a machine onto a traditional loom and automate the weaving of complex patterns in fabric, more complex than could typically be handled manually. 

Jacquard became the father of automation. He received praise from Napoleon and his machine was of great interest to the brilliant mind of Ada Lovelace (read her bio and impact on AI).

The machine was an important step on the road to the industrial revolution. But the canuts were worried about losing their jobs. 

Go cylindrical and multiply. In 1673, Leibniz invented a cylinder lined with 9 teeth of varying lengths that would couple with a counting wheel placed at a certain height with respect to the cylinder. At a medium height, it might ecounter 4 teeth for every turn of the cylinder and advance by 4 steps, to simulate the addition of 4 to a current total. 

In 1694, the Stepped Reckoner machine used those cylinders to automate all four arithmetic operation, paving the way for all mechanical calculators of the next 3 centuries.

Much like the seesaw at the very top of this page, this would no longer be considered to be Artificial Intelligence today. But in 1642, the Pascaline presented a groundbreaking way for Blaise Pascal to calculate the taxes of the French region Haute Normandie on behalf of cardinal de Richelieu.

Via spoked metallic wheels, Pascal's calculator automated two arithmetic operations (addition and subtraction) and enabled others. Pascal received Royal Priviledge allowing him to sell copies of the machine, but costs cut the commercial career short. But it layed the groundwork for other designs to come in the future.

In Greek mythology, the God Hepaestus was said to have built a bronze automaton to protect Crete, at the request of Zeus. It was designed to perform tasks without human intervention and appears to be the first record of our fascination for artificial beings. Among feats described in the Argonautica, it hurled giant boulders at the Argo to keep it at bay from Crete. 

Since our photographic records of the times are scarce, I asked Midjourney for a representation. This image was one of the four the system gave me in return for the simple prompt. Yes, that's a camera, on the right. After all, a 30 m tall bronze robot performing war feats worthy of legend isn't going do all that without posting on the Gram, right? 😉

Pandora, described today as a woman, was also reported to be an artificial being by Hesiod in 700 BC, along with other beings such as mechanical servants made of gold (Source).