By the beginning of the twenty-first century, computers had become extraordinarily powerful machines.
They could process enormous quantities of information, perform billions of calculations, store vast databases and communicate almost instantly through global networks. The internet had also created something earlier generations of computer scientists had never possessed at comparable scale: an expanding ocean of digital data produced by governments, businesses, scientific instruments and billions of ordinary people.
Yet most computers still operated according to a familiar principle. Human beings wrote instructions and the computer followed them. If a programmer wanted a machine to calculate tax, process an account or sort a database, the programmer described the steps necessary to perform that task. Computers could execute those instructions with remarkable speed and consistency, but the intelligence behind the procedure largely remained with the person who had written the rules.
Artificial intelligence introduced a different ambition.
Instead of telling a computer exactly how to solve every problem, what if researchers could show it examples and allow the machine to discover useful patterns for itself? What if a computer could become better at a task through experience?
That question produced one of the most important transformations in the history of computing.
The machine did not literally begin learning in the way a child learns, nor did it suddenly acquire consciousness or human understanding. What changed was more precise. Researchers developed mathematical systems capable of adjusting their internal parameters in response to data, allowing computers to improve at tasks without a programmer explicitly specifying every decision they should make.
The machine was no longer limited to following rules written entirely in advance. It could begin finding patterns.
The Dream of Intelligent Machines Came Before the Technology
The ambition to create intelligent machines is almost as old as modern computing itself.
In 1950, Alan Turing published Computing Machinery and Intelligence and opened with the question that would come to define much of the field: whether machines could think. Rather than becoming trapped in arguments over the precise definition of thought, Turing proposed examining whether machine behaviour could become sufficiently sophisticated to resemble intelligent human behaviour. His paper also considered the possibility of learning machines rather than systems whose behaviour had been completely fixed by their designers.
Six years later, a small group of researchers gathered at Dartmouth College for the Dartmouth Summer Research Project on Artificial Intelligence. John McCarthy and the other organisers approached the subject with extraordinary confidence, proposing that aspects of learning and intelligence might be described precisely enough for machines to simulate them.
The meeting helped establish artificial intelligence as a recognisable field of research and gave it the name that remains with us today.
The ambition was enormous. Researchers wanted computers to reason, solve problems, understand language, recognise objects and perform tasks that appeared to require intelligence. What they did not yet know was which path would lead there.
One influential approach treated intelligence primarily as a problem of knowledge and reasoning. If experts knew how to solve a problem, perhaps their knowledge could be translated into rules that a computer could follow.
A medical expert, for example, might reason that a particular combination of symptoms, laboratory results and patient history suggested one diagnosis rather than another. If those relationships could be written as rules, a computer could apply them consistently.
This approach became particularly important in the development of expert systems. Stanford’s DENDRAL project, begun in the 1960s, applied specialist knowledge to problems in chemistry, while the later MYCIN system used a substantial collection of rules to advise on certain bacterial infections and antibiotic treatments.
Such systems demonstrated that computers could perform impressively in narrow domains when supplied with carefully structured expert knowledge. But the approach contained a weakness.
Someone had to write the knowledge down.
Every new rule had to be discovered, articulated and entered into the system, while real life contained ambiguities and exceptions that were difficult to express through rigid logical statements.
Human beings often recognise a face without being able to describe every rule that distinguishes that face from millions of others. A doctor may notice a pattern that experience has made familiar without reducing the judgement to a neat sequence of instructions. A driver can identify a pedestrian under very different lighting conditions without consciously measuring every relevant visual feature.
Some problems therefore seemed to require something other than increasingly elaborate rule books. The machine needed a way to learn from examples.
One of the earliest demonstrations came from an unexpected place. A board game.
Arthur Samuel began developing a checkers program at IBM during the 1950s. The machine could not simply calculate every possible future move because even a relatively constrained game generates too many possibilities for exhaustive search on the computers of that era. Samuel instead developed techniques that allowed the program to evaluate positions and improve its play using experience from previous games.
The important achievement was not simply that a computer could play checkers. Computers had already demonstrated that they could follow game rules. The more consequential idea was that experience from earlier games could alter future performance.
Samuel called this approach machine learning, helping establish the term that would later define one of the most important branches of artificial intelligence.
At roughly the same time, Frank Rosenblatt was developing the perceptron, a system inspired loosely by networks of neurons in the brain. Its internal connections had adjustable strengths, and when the system made an incorrect prediction, those weights could be modified so that future predictions became more accurate.
In 1958, Rosenblatt and his colleagues built the Mark I Perceptron, an electronic implementation capable of learning to distinguish simple patterns rather than requiring every classification rule to be explicitly programmed beforehand.
Its abilities were extremely limited by modern standards, but the underlying concept was profound. A system could make a prediction, compare the prediction with the desired result, adjust itself and gradually improve through repeated examples. The idea was simple, but making it work at useful scale would take decades.
Artificial Intelligence Had to Survive Its Own Expectations
Early AI research generated enormous optimism.
Researchers produced programs capable of solving mathematical problems, playing games, manipulating symbols and performing impressive demonstrations inside carefully controlled environments. It became tempting to assume that broadly intelligent machines were close. They were not.
Computers remained expensive and relatively weak. Digital data was scarce by modern standards, and many algorithms that worked impressively on small laboratory problems struggled when confronted with the complexity and ambiguity of the real world.
Rule based AI encountered difficulties when knowledge bases became too large and brittle. Early neural networks also had significant mathematical and computational limitations, while expectations repeatedly moved ahead of what the technology could actually deliver.
The result was not the disappearance of AI research, but periods in which enthusiasm and funding contracted. These became known as AI winters. Important research continued throughout them, but the pattern was clear: extravagant expectations were followed by disappointment when promised capabilities failed to arrive.
That history matters because today’s progress can make the development of artificial intelligence appear inevitable. It was anything but inevitable. For decades, researchers repeatedly encountered the same frustrating reality: the idea was often ahead of the machinery needed to realise it.
Even while public enthusiasm rose and fell, machine learning continued developing. Researchers increasingly approached intelligence as a statistical problem rather than attempting to construct complete symbolic descriptions of the world.
Suppose a bank wants to determine whether a transaction is suspicious. Writing rules might work for obvious cases, but fraud patterns change, thousands of variables may matter and the relationships between them may be difficult to specify manually.
A learning system can instead examine large collections of previous transactions, compare fraudulent examples with legitimate ones and identify combinations of features associated with unusual behaviour.
The same principle can be applied to spam filtering, credit assessment, handwriting recognition, product recommendations and many other problems. These systems do not need to understand fraud, handwriting or preference in the human sense. They need to identify statistical relationships useful enough to make predictions.
That distinction became central to machine learning.
Neural network research was also acquiring better tools. Influential work by David Rumelhart, Geoffrey Hinton and Ronald Williams published in 1986 helped popularise the use of backpropagation for training multilayer neural networks.
The technical mathematics can become complicated, but the basic logic is intuitive. A network makes a prediction, its error is measured, and information about that error is propagated backwards through the system so that its internal weights can be adjusted. Repeated across many examples, those adjustments can gradually improve performance.
This allowed hidden layers inside neural networks to discover useful intermediate representations that programmers had not explicitly specified. The theory was improving, but the surrounding technological world still needed to catch up.
Data and Computing Finally Caught Up With the Idea
The breakthrough did not come from machine learning alone. It came because several technological developments that had been unfolding separately began reinforcing one another.
The personal computer had spread computation through homes and offices. The internet connected machines globally. Businesses moved records into databases. Digital cameras transformed photographs into data. Search engines processed billions of questions. Online stores recorded purchasing behaviour. Mobile phones became sensors carried throughout ordinary life, while social platforms accumulated photographs, relationships and interactions on an unprecedented scale.
Human activity was becoming machine readable.
At the same time, semiconductor technology continued making computation dramatically more powerful, digital storage became cheaper and global computing infrastructure became increasingly accessible.
By the early twenty first century, machine learning researchers possessed something their predecessors had lacked: enormous datasets, powerful processors, affordable storage, improved algorithms and increasingly sophisticated computing infrastructure.
An especially important development came from hardware that had originally been designed for computer graphics.
Graphics processing units perform enormous numbers of similar mathematical operations in parallel. Researchers realised that many of the calculations required to train neural networks had similar characteristics, allowing GPUs to accelerate training dramatically.
This did not invent neural networks or machine learning. What it changed was scale.
Networks containing far more parameters could now be trained on far larger datasets within practical amounts of time. The technology was finally becoming powerful enough for ideas that had existed for decades to produce unexpectedly capable results.
One of the clearest demonstrations arrived in 2012.
The ImageNet project had assembled a very large labelled collection of images covering thousands of categories, allowing researchers to compare systems attempting to recognise objects in photographs.
That year, Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton entered a deep convolutional neural network later widely known as AlexNet. The system was trained on roughly 1.3 million high resolution images covering 1,000 categories and used GPUs to accelerate the workload.
Its performance was dramatically better than competing approaches. The AlexNet paper reported a top five error rate of 15.3 per cent compared with 26.2 per cent for the second best system in the competition.
The importance of the result extended well beyond recognising photographs. It demonstrated what could happen when neural networks, enormous datasets and powerful parallel computation were combined.
Deep learning was no longer simply an interesting academic approach. It was becoming an engineering advantage.
Traditional computer vision had often depended heavily on human engineers deciding which visual features software should examine. Edges might matter, corners might matter, shapes might matter and textures might matter, so engineers designed methods for extracting those characteristics before passing them to a classifier.
Deep neural networks changed part of that process.
Given enough examples, successive layers of a network could learn increasingly complex representations. An early layer might become sensitive to simple edges or colour transitions, while later layers could combine those signals into textures and shapes. Still later layers could develop representations useful for distinguishing a dog from a car or a face from a building.
The programmer still designed the learning system, chose the training process and supplied the data, but no longer had to hand write every useful visual feature. The machine could discover representations through training.
Prediction Quietly Became Part of Everyday Life
Long before artificial intelligence became a daily public conversation, machine learning was already entering ordinary life.
Email providers used statistical systems to identify spam. Banks used models to detect unusual transactions. Search engines learned better ways to rank information. Online retailers analysed behaviour to recommend products. Streaming platforms estimated what viewers might want to watch, phones improved speech recognition, cameras detected faces and digital mapping services used large datasets to estimate travel conditions.
Hospitals and research laboratories also began experimenting with pattern recognition in medical information.
Most people interacted with these systems without describing the experience as artificial intelligence.
There was no humanoid robot and no dramatic moment when a machine announced that it had become intelligent. Machine learning appeared instead as a recommendation, a search result, a fraud warning, a photograph automatically organised into an album, a voice converted into text or a suspicious email quietly moved into a spam folder.
The revolution was already underway before most people noticed it.
What machine learning increasingly automated was not simply physical activity but prediction.
A conventional automated system can follow a predetermined sequence repeatedly. A machine learning system can examine data and estimate what is likely to happen next or determine which category a new example most probably belongs to.
Is this transaction fraudulent? Does this image show a particular abnormality? Which advertisement is this user most likely to click? Which customer is most likely to cancel a subscription? Which route will probably take the least time? Which product resembles others this customer previously bought?
These are prediction problems.
Once prediction becomes cheaper and more widely available, decisions that previously required extensive human judgement can begin incorporating machine generated probabilities. The machine does not necessarily make the final decision, but it increasingly influences which decision a human being sees as reasonable.
That is a different kind of automation from the factory robot. It enters judgement.
Not every learning system requires a teacher to provide the correct answer for every example either. Another important branch of machine learning became known as reinforcement learning, in which an agent performs an action, receives some form of reward or penalty from its environment and gradually learns which behaviours tend to produce better outcomes.
Games became particularly useful laboratories for these ideas because the rules and objectives are clearly defined, winning provides a measurable goal and millions of simulated experiences can sometimes occur far faster than equivalent experimentation in the physical world.
This approach contributed to one of the most dramatic public demonstrations in the history of artificial intelligence.
The ancient board game Go had long posed a formidable challenge for computers because its enormous number of possible positions made straightforward brute force approaches exceptionally difficult.
DeepMind’s AlphaGo combined deep neural networks with search techniques, learning initially from examples of expert human play before improving further through games played against versions of itself.
In 2015 it defeated European champion Fan Hui, and in March 2016 it faced Lee Sedol, one of the world’s strongest Go players, in Seoul.
AlphaGo won four games to one.
The victory attracted global attention because Go had long been regarded as a domain requiring intuition as well as calculation. What the system demonstrated was more precise, however. Learning systems could discover highly effective strategies in problem spaces too large for programmers to describe exhaustively.
The machine was no longer winning only because engineers had told it exactly which moves were good. Training had shaped its ability to evaluate the game.
Learning Changed the Machine Without Making It Human
The achievements were striking, but they encouraged a dangerous linguistic shortcut.
When we say that a machine learns, it is easy to imagine something much closer to human learning than what is technically occurring.
A child can learn that a dog is an animal while simultaneously developing ideas about movement, life, emotion, danger, friendship and the broader world in which dogs exist. A machine learning classifier may distinguish dogs from cats with extraordinary accuracy while possessing none of that surrounding understanding.
It has learned a statistical relationship useful for classification. That does not automatically mean it possesses human concepts.
The distinction becomes particularly important because machine learning systems can appear highly capable while remaining fragile. A model learns from the information available to it, which means that distortions in the data can become distortions in the model. If important populations are poorly represented, performance can vary between groups. If real world conditions change significantly from the circumstances represented during training, accuracy can deteriorate.
A model optimised for one objective may also behave unexpectedly when inserted into a broader social environment.
Machine learning therefore did not eliminate human responsibility. It changed where that responsibility had to be exercised.
Choosing the data matters. Choosing the objective matters. Evaluating performance matters. Deciding where the system should be used matters. Understanding its limitations matters.
The rise of machine learning also transformed the economic importance of data.
During the earlier computer revolution, data was valuable because computers could store and process it. Machine learning made large collections of data valuable for another reason: they could be used to train models.
A company with millions of customer interactions could identify behavioural patterns. A search engine could use enormous records of queries and clicks to improve ranking. A mapping service could analyse movements from connected devices to estimate traffic. A retailer could learn relationships between products and buyers, while medical researchers could search datasets for patterns difficult for individual investigators to detect.
Data was becoming more than a record of what had happened. It could help predict what might happen next.
Look back across the history we have followed in this series and a larger pattern begins to emerge.
The steam engine multiplied physical force. Electricity and mass production multiplied industrial scale. The computer turned one physical machine into many possible machines through software. Digitisation turned information into something those machines could process. The internet connected them and generated enormous quantities of new data.
Machine learning added another layer by allowing computers to modify aspects of their behaviour according to patterns discovered within that data.
The difference may sound subtle, but its consequences were enormous.
A programmer no longer needed to anticipate every possible example. Instead, the programmer could construct a learning process, and some of the relationships governing the machine’s final decisions could emerge through training.
That transformed what computers could reasonably attempt. Images, speech, medical records, financial behaviour, consumer preferences, scientific measurements and complex games could all become domains in which machines searched for statistical structure.
For most of the history described here, however, machine learning was principally concerned with recognising, classifying, ranking and predicting. A system might identify what appeared in an image, determine whether a transaction looked suspicious, recommend a product, choose a move in a game or convert speech into text.
The computer was becoming remarkably good at detecting patterns in information created by human beings.
Then researchers began asking a different question.
Instead of asking only what a machine could recognise in existing information, what if it could use the patterns it had learned to produce something new? Could a machine create an image that had never existed? Could it produce convincing human language? Could it write software, compose music or generate video?
That would mark another change in the relationship between people and machines. But it belongs to the next chapter.
For now, the crucial transformation is simpler.
Humanity had spent centuries building machines and telling them exactly what to do. Machine learning introduced a different possibility. We could build a machine, expose it to experience and allow part of its capability to emerge from the data itself.
Machines had learned to learn.
Next in The Trend Report
Machines Learned to Create
Generative AI Changed What Computers Could Produce
Resources
This article draws principally on Alan Turing’s 1950 paper Computing Machinery and Intelligence published in Mind; Dartmouth College’s historical archive of the 1956 Dartmouth Summer Research Project on Artificial Intelligence; IBM’s archival history of Arthur Samuel’s checkers program and the emergence of the term machine learning; the Smithsonian National Museum of American History and Cornell University archives on Frank Rosenblatt and the Mark I Perceptron; Stanford and Computer History Museum material on expert systems; David Rumelhart, Geoffrey Hinton and Ronald Williams’ 1986 work on backpropagation; the 2012 AlexNet paper by Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton; and the Nature publications documenting deep learning and AlphaGo.
One Response
‘A model learns from the information available to it, which means that distortions in the data can become distortions in the model. If important populations are poorly represented, performance can vary between groups. If real world conditions change significantly from the circumstances represented during training, accuracy can deteriorate.’… Bingo!!
If these machines learn from data fed into them, they are not that intelligent anyway. They cannot experience something new, they only modify what they already have in more ‘intelligent’ ways.. 🙌