Thinking Machines: AI from 1843 to 2011

Before computers: the dream of a thinking machine

People imagined thinking machines long before anyone could build one: the bronze giant Talos in Greek myth, clockwork automata that wrote letters and played music in the 1700s. The modern story starts when the idea met mathematics.

1950 to 1956: a field gets its name

In 1950 Alan Turing published Computing Machinery and Intelligence. Instead of arguing about what "thinking" means, he proposed a test he called the imitation game: a judge holds a typed conversation with a hidden human and a hidden machine, and if the judge cannot reliably tell which is which, the machine has done something we would call intelligent. It became known as the Turing test. Turing predicted that by about the year 2000 machines would fool an average judge in five minutes about 30% of the time.

Programs that played games and proved theorems followed quickly. Christopher Strachey wrote a draughts program for the Ferranti Mark 1 in Manchester in 1951. In 1956 the Logic Theorist, by Allen Newell, Herbert Simon and Cliff Shaw, proved 38 of the first 52 theorems in Whitehead and Russell's Principia Mathematica.

The Dartmouth workshop, 1956

In 1955 four researchers, John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, proposed a summer workshop at Dartmouth College. Their proposal used the phrase artificial intelligence for the first time and set out the field's founding bet:

The workshop ran in the summer of 1956. It did not solve intelligence, but it gave the field a name, a community and an agenda, which is why 1956 is usually given as the year AI was born.

The golden years and the first AI winter

The next fifteen years were full of confidence.

• 1958: Frank Rosenblatt demonstrated the perceptron, a machine that learned to recognise simple patterns from examples using 400 light sensors as its eye. A newspaper reported it as the embryo of a computer that would walk, talk and be conscious.
• 1959: Arthur Samuel coined the term machine learning for his draughts program, which improved by playing against itself until it could beat him.
• 1964 to 1966: Joseph Weizenbaum's ELIZA held conversations by matching patterns in what people typed and turning them into questions. Many users felt it understood them, even though it understood nothing. That reaction is still called the ELIZA effect.
• 1966 to 1972: Shakey, at the Stanford Research Institute, became the first mobile robot that could reason about its own actions.

In 1965 Herbert Simon predicted that "machines will be capable, within twenty years, of doing any work a man can do". Progress was real, but it was nowhere near that fast.

Why the winter came

• In 1969 Minsky and Papert's book Perceptrons proved that a single-layer perceptron cannot learn simple functions such as XOR. Interest in neural networks collapsed for over a decade.
• Programs that worked on toy problems failed on real ones: the number of possibilities exploded, and computers were far too slow and small.
• In 1973 the Lighthill Report told the UK government that AI had failed to deliver on its promises, and funding was cut. American funding was redirected too.

The years from about 1974 to 1980 are called the first AI winter: less money, fewer students, and a reputation for overpromising.

Expert systems, a second winter, and learning machines

The expert systems boom

In the 1980s AI found money again with expert systems: programs that stored a specialist's knowledge as hundreds or thousands of IF-THEN rules.

Japan launched its Fifth Generation computer project in 1982, and other governments responded with programmes of their own.

But rules were expensive to write, brittle when the world changed, and impossible to keep consistent at large scale. When the market for specialised AI computers collapsed in 1987, the second AI winter followed, lasting into the early 1990s.

Meanwhile, learning was quietly coming back

• 1986: David Rumelhart, Geoffrey Hinton and Ronald Williams showed how backpropagation can train networks with hidden layers, the answer to the problem Perceptrons had exposed.
• 1989: Yann LeCun trained a convolutional neural network to read handwritten postcodes. Its descendants were reading a large share of cheques in American banks by the late 1990s.
• 1997: IBM's Deep Blue beat world chess champion Garry Kasparov. It searched about 200 million positions a second with rules written by experts: impressive, but not learning.
• 1997: Sepp Hochreiter and Jürgen Schmidhuber invented the LSTM, a network with memory for sequences.
• 1990s and 2000s: statistical machine learning powered spam filters, search engines and recommendations without anyone calling it AI.
• 2006: Hinton and colleagues showed how to train deep networks layer by layer, and the term deep learning caught on.
• 2009: ImageNet, led by Fei-Fei Li, labelled over 14 million images, giving vision researchers the data deep learning would need.
• 2011: IBM's Watson won the quiz show Jeopardy!, and Apple's Siri put a voice assistant in millions of pockets.

Hands-on: rules versus learning

Build ELIZA in ten lines

ELIZA had no understanding at all: it matched patterns and reflected your words back as questions. This version follows the same idea. Run it, then add your own rules.

import re

RULES = [
    (r"I need (.*)", "Why do you need {0}?"),
    (r"I am (.*)", "How long have you been {0}?"),
    (r"I feel (.*)", "What makes you feel {0}?"),
    (r"(.*) mother(.*)", "Tell me more about your family."),
    (r"(.*)\?", "Why do you ask that?"),
]

REFLECT = {"i": "you", "me": "you", "my": "your", "am": "are",
           "you": "I", "your": "my"}

def reflect(words):
    # Turn "my exams" into "your exams", as ELIZA did
    return " ".join(REFLECT.get(w.lower(), w) for w in words.split())

def eliza(sentence):
    for pattern, reply in RULES:
        match = re.match(pattern, sentence, re.IGNORECASE)
        if match:
            return reply.format(*[reflect(g) for g in match.groups()])
    return "Please, go on."

for line in ["I need a holiday", "I am worried about my exams",
             "I feel tired", "My mother says I work too hard", "Is this real?",
             "The weather is nice"]:
    print("You:  ", line)
    print("ELIZA:", eliza(line))

Every reply was written by a person in advance. Nothing was learned, and nothing new could ever be said. That is symbolic AI: intelligence written down as rules. The rest of this course is about the other approach, where the rules are learned from data.

Activity: the imitation game

Work in threes: a judge, a human and a "machine". The human answers the judge's written questions honestly. The machine may only answer using a list of ten stock phrases agreed in advance. The judge has five questions to decide which is which. Then discuss: what questions gave the machine away, and what would a machine need to pass?

Activity: timeline challenge

Write each event from this lesson on a card, shuffle them, and put them in order without looking. Then sort them into three piles: ideas, rules and symbols, and learning. Which pile grows fastest after 1986?

Level check

Before moving on, make sure you can:

• Explain what Turing's imitation game tests, and why he proposed it
• Say why 1956 is called the birth of AI, and name two people involved
• Describe what a perceptron was, and what Minsky and Papert proved about it
• Explain what an AI winter is, and give two causes of the first one
• Compare an expert system with a learning system
• Name three milestones between 1986 and 2011 that made deep learning possible

Key terms you must be able to define

Artificial intelligence

The field of making machines do things that would need intelligence if a person did them, a name first used in the 1955 Dartmouth proposal.

Turing test

Turing's imitation game: a machine passes if a judge cannot reliably tell it from a human in typed conversation.

Perceptron

Rosenblatt's 1958 learning machine: a single artificial neuron that adjusts its weights from examples.

Machine learning

Programs that improve at a task from data or experience rather than from rules written by a person. Coined by Arthur Samuel in 1959.

Symbolic AI

AI built from explicit rules and symbols written by people, such as ELIZA and expert systems.

Expert system

A program storing a specialist's knowledge as many IF-THEN rules.

AI winter

A period of reduced funding and interest after AI failed to meet its promises: roughly 1974 to 1980, and 1987 to the early 1990s.

ELIZA effect

The tendency to believe a program understands us when it is only matching patterns.

Backpropagation

The method for training networks with hidden layers by passing the error backwards, popularised in 1986.

ImageNet

A dataset of over 14 million labelled images, released in 2009, that fuelled the deep learning breakthrough.

Flashcards

Go further: notes for future AI engineers

Read the originals

Turing's 1950 paper Computing Machinery and Intelligence is short and readable. Its sixth section answers nine objections to thinking machines, including one he called "Lady Lovelace's Objection". The 1955 Dartmouth proposal is only a few pages long. Reading the founders' own words shows how much they got right, such as learning, language and abstraction, and how badly they underestimated the time it would take.

Symbolic versus connectionist AI

The history of AI is a long argument between two camps. Symbolic AI, nicknamed "good old-fashioned AI", holds that intelligence is the manipulation of symbols by rules. Connectionism holds that it emerges from many simple units learning the strengths of their connections. Symbolic AI dominated until the late 1980s; connectionism, as deep learning, dominates now. Many researchers expect the future to combine both, with neural networks that can also search, use tools and reason explicitly, a direction called neuro-symbolic AI.

Misconceptions to avoid