Machine learning

• 1. Introduction

  ›

  Machine learning is the field that gave us a computer program capable of beating the world's top players at Go, a game long considered too intuitive for machines to master. That milestone arrived in 2016, when AlphaGo used reinforcement learning techniques to defeat those champions. It punctuated a story that stretches back decades, to an IBM researcher named Arthur Samuel who coined the term "machine learning" in 1959. The questions this documentary will answer are deceptively simple: How did a cluster of theoretical ideas about the human brain become algorithms that filter your email, flag fraudulent bank transactions, and help doctors diagnose disease? And what happens when those systems go wrong?

• 2. Roots in Human Cognition

  ›

  In 1949, Canadian psychologist Donald Hebb published The Organization of Behavior, a book that proposed a theoretical neural structure built from interactions among nerve cells. Hebb's model of how neurons strengthen their connections through repeated activation became a foundation for how many machine learning algorithms are designed today. Walter Pitts and Warren McCulloch extended this line of thinking by proposing the first mathematical model of neural networks, with algorithms meant to mirror human thought processes.

  Arthur Samuel's earliest machine learning program, introduced in the 1950s, calculated the probability of winning at checkers for each side. Samuel himself used the synonym "self-teaching computers" for what he was building. The ambition at that time was straightforwardly cognitive: could machines replicate human reasoning? Alan Turing had posed a version of this question in his paper "Computing Machinery and Intelligence," replacing "Can machines think?" with the more testable question of whether machines could convincingly imitate a human in conversation.

  By the early 1960s, the Raytheon Company had developed an experimental learning machine called Cybertron, which used punched tape memory to analyse sonar signals, electrocardiograms, and speech patterns. Cybertron was trained by a human operator who could press a "goof" button to signal incorrect decisions. Nils Nilsson's book "Learning Machines," published during that decade, captured the era's focus on pattern classification. Interest in pattern recognition continued into the 1970s, and by 1981 researchers were reporting systems capable of recognising 40 characters from a computer terminal, including 26 letters, 10 digits, and 4 special symbols.

• 3. The Split From Artificial Intelligence

  ›

  By 1980, expert systems had come to dominate the field of artificial intelligence, and statistical approaches were out of favour. The divide between AI and machine learning was real: probabilistic systems were plagued by problems in data acquisition and representation, while AI researchers pressed toward logical, knowledge-based methods. Neural network research was abandoned by both AI and computer science communities around the same time.

  A group of researchers from outside those core disciplines kept the work alive. John Hopfield, David Rumelhart, and Geoffrey Hinton continued what was called "connectionism," and their most significant success came in the mid-1980s with the reinvention of backpropagation, the technique for adjusting a network's internal connections based on its errors. Machine learning then reorganised itself as its own field in the 1990s, deliberately shifting its goal away from achieving artificial general intelligence and toward solving practical, tractable problems. It borrowed methods from statistics, fuzzy logic, and probability theory rather than the symbolic approaches inherited from AI.

  Tom M. Mitchell gave the field a widely quoted formal definition: "A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P if its performance at tasks in T, as measured by P, improves with experience E." That definition is deliberately operational. It describes what a learning system does, not what it is, placing the emphasis on measurable improvement over time.

• 4. How Machines Actually Learn

  ›

  Machine learning approaches fall into three broad categories. In supervised learning, a program is shown example inputs paired with their correct outputs, supplied by a teacher, and its goal is to learn the general rule that connects them. Classification algorithms assign inputs to categories; a spam filter is a classic example, where an incoming email is the input and the folder it belongs in is the output. Regression algorithms handle tasks where the output is a continuous number, such as forecasting temperatures from historical data.

  Unsupervised learning operates without labels. The algorithm searches for structure in raw data on its own, identifying clusters and patterns without being told what to look for. A special variant called self-supervised learning, introduced in 1982 alongside a neural network called the crossbar adaptive array, generates its own training signals from the data itself. That system replaced external rewards with an internal model of emotion as a state evaluation mechanism, driven by the interaction between cognition and emotion in the algorithm's design.

  Reinforcement learning places a program in a dynamic environment where it must achieve a goal and receives rewards as feedback. The program learns by trying to maximise cumulative reward over time. Autonomous vehicles and game-playing programs are two of its most cited applications. No single algorithm works for all problems; the appropriate category depends on what kind of signal or feedback is available.

• 5. Deep Learning and the Hardware Race

  ›

  In 2014, Ian Goodfellow and others introduced generative adversarial networks, or GANs, which could produce realistic synthetic data. Deep learning, which stacks multiple hidden layers inside an artificial neural network, had by then already produced major results in computer vision and speech recognition. The approach models the way the human brain processes light and sound into perception.

  The hardware demands of deep learning are staggering. OpenAI tracked the hardware compute used in the largest deep learning projects from AlexNet in 2012 to AlphaZero in 2017 and found a 300,000-fold increase in the amount of compute required, with a doubling time of 3.4 months. By 2019, graphics processing units with AI-specific enhancements had displaced CPUs as the dominant method for training large-scale commercial cloud AI.

  Google responded to these demands by developing Tensor Processing Units, or TPUs, specialised hardware accelerators optimised for tensor computations. Introduced in 2016, TPUs are built around matrix multiplication units and high-bandwidth memory and are used to power models including DeepMind AlphaFold. A different approach, neuromorphic computing, attempts to emulate biological neural networks in hardware, using electrically adjustable materials such as memristors to replicate the function of neural synapses rather than simulating them in software.

• 6. Bias, Failures, and the Ethics Gap

  ›

  In 1988, the UK's Commission for Racial Equality found that St. George's Medical School had been using a computer program trained on historical admissions data. That program had denied nearly 60 candidates who were women or had non-European-sounding names. The case is an early example of a pattern that has repeated across many domains: machine learning systems trained on biased human-generated data absorb and reproduce those biases.

  In 2015, Google Photos tagged images of Black people as gorillas. In 2016, Microsoft tested a chatbot called Tay that learned from Twitter and rapidly acquired racist and sexist language. An investigation by ProPublica found that a recidivism prediction algorithm flagged Black defendants as high risk twice as often as white defendants. Geolitica's predictive policing algorithm produced disproportionately high levels of over-policing in low-income and minority communities after being trained on historical crime data.

  Research by the Computing Research Association in 2021 found that female faculty made up just 16.1% of AI faculty at universities surveyed. Among new U.S. resident AI PhD graduates, 45% identified as white, 22.4% as Asian, 3.2% as Hispanic, and 2.4% as African American. Fei-Fei Li, one of the scientists raising concerns about fairness, put the stakes plainly: "There's nothing artificial about AI. It's inspired by people, it's created by people, and most importantly it impacts people."

  Failures have not been limited to bias. In 2018, a self-driving car from Uber failed to detect a pedestrian, who was killed. IBM's Watson system, applied to healthcare after years of investment and billions of dollars spent, failed to deliver the promised results. These outcomes reflect what critics call the black box problem: even the engineers who build a machine learning system often cannot audit the pattern it extracted from the data or explain a specific decision it made.

• 7. Reach and Consequence

  ›

  In 2006, Netflix launched a competition offering a grand prize of one million dollars to any team that could improve its existing Cinematch movie recommendation algorithm by at least 10%. A joint team from AT&T Labs-Research, collaborating with two groups called Big Chaos and Pragmatic Theory, won that prize in 2009 with an ensemble model. Shortly afterward, Netflix concluded that viewers' star ratings were not actually the best signal for predicting what they would watch and changed its recommendation engine based on that insight.

  In 2010, Rebellion Research used machine learning to analyse the 2008 financial crisis, as reported in The Wall Street Journal. In 2012, Vinod Khosla, co-founder of Sun Microsystems, predicted that 80% of medical doctors' jobs would be lost within two decades to machine learning diagnostic software. In 2019, Springer Nature published the first research book created using machine learning. By 2020, machine learning was being applied to help make diagnoses and assist researchers developing a response to COVID-19.

  Federated learning represents one answer to the privacy concerns raised by systems that require centralised data. It trains models on users' devices without sending individual data to a server. Google's Gboard keyboard uses federated machine learning to train search query prediction models directly on mobile phones. Springer Nature's 2019 machine learning book and Gboard's on-device training point toward a field increasingly aware that the reach of its systems demands accountability alongside capability.