# What did Alan Turing propose in his 1948 report 'Intelligent Machinery'?

Date: 2019-11-22 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-was-turings-intelligent-machinery-report/

**Answer:** In "Intelligent Machinery," a 1948 report for Britain's National Physical Laboratory, Alan Turing argued that machines could be made to behave intelligently and that the best route was to educate them, starting from randomly wired "unorganized machines" that a teacher shapes with rewards and punishments. The report stayed unpublished until 1969, yet it contains early versions of neural networks, learning by reward and evolutionary search.

Two years before his famous paper on the imitation game, Alan Turing wrote a report that asks a bolder question in plainer terms: can a machine show intelligent behavior, and if so, how would we build one? His answer was that we should not try to program a finished mind. We should build something closer to a baby's brain and teach it.

The report, titled "Intelligent Machinery," was written for the National Physical Laboratory (NPL) in Teddington, near London, where Turing had been designing an electronic computer called the ACE. It is dated 1948. His employers did not publish it, and it first appeared in print in 1969, in a book series called *Machine Intelligence*. That long delay is one reason many readers in 2019 still meet Turing only through the Turing test.

## What the report says

Turing opens by listing why people resist the idea of thinking machines. Some dislike the thought of a rival to human intellect. Some see it as irreverent. Some judge all machinery by the simple, repetitive devices in use until about 1940. Some point to mathematical results showing that any fixed machine will fail on certain questions. And some say a machine's cleverness is only its maker's cleverness in disguise.

He answers each in turn. A machine is allowed to make mistakes, he notes, and so are clever children. A pupil's discoveries are not credited to the teacher, so why should a machine's be credited to its builder?

Then comes the heart of the report. Turing describes what he calls "unorganized machines": large numbers of identical simple units wired together more or less at random. Each unit has two inputs and one output, and a central clock ticks them all forward together. At every tick, each unit looks at its two inputs and produces a 0 or a 1 by a fixed rule. Turing calls this the "A-type" machine and describes it as roughly the most basic model you could make of a nervous system whose neurons are wired at random.

He also defines a "B-type" machine, in which each connection passes through a small switch that can be set from outside. By setting those switches, an operator can turn a random tangle into a machine that does a useful job. Turing calls this setting-up process "organizing" the machine, and he compares it with education.

## Learning by reward and punishment

How would a teacher organize such a machine without opening it up? Turing's suggestion is to give it just two extra inputs: one for "pleasure" (reward) and one for "pain" (punishment). When the machine does something right, a reward makes its recent choices permanent. When it does something wrong, a punishment wipes those tentative choices so it can try again.

He worked through a version he called a P-type machine entirely by hand, on paper, because no suitable computer was free. He kept its tentative choices in pencil and its permanent ones in ink. In one run, the machine fell into a loop, repeating the same two actions; a single punishment broke the loop. Turing reported that, with an added external memory, he had succeeded in training one of these paper machines to become a universal machine, meaning one that can imitate any other computing machine, though he admitted the method was not much like real teaching.

Here is the same idea as a tiny classroom exercise you can do with a coin:

| Step | What you do | What it models |
|---|---|---|
| 1 | Write three blank boxes: Left, Right, Stop | An untrained machine with no fixed habits |
| 2 | When a box is blank, flip a coin and pencil in the result | A random, tentative choice |
| 3 | If the outcome is good, ink over the pencil | Reward makes the choice permanent |
| 4 | If the outcome is bad, erase all pencil marks | Punishment sends it back to random |

> **Try it:** Run the coin exercise above with a friend acting as "teacher" who secretly wants the sequence Left, Left, Stop. Count how many rounds it takes before the inked boxes match the goal.

## The brain as an untrained machine

Turing went further and suggested that the cortex of a newborn child is itself an unorganized machine. Parts of the brain handle fixed jobs such as breathing, he wrote, but the parts used for thinking are shaped by training. His example was simple: children on the two sides of the English Channel grow up speaking different languages because of training, not because their brains are built differently.

He also proposed fields where a "brain" without a body could practice: games such as chess, learning and translating languages, codebreaking and mathematics. And he named three kinds of search that might explain intelligence: intellectual search inside one mind, genetic or evolutionary search across generations, and cultural search carried out by a whole community.

## Why the report matters

The Stanford Encyclopedia of Philosophy notes that Turing's trainable nets of logical components foreshadow neural networks, and that his "genetical or evolutionary search" anticipates what are now called genetic algorithms. Because the report sat in a file for two decades, though, it had little direct influence.

Seen from late 2019, the report reads as surprisingly current. Learning from rewards is the core idea behind reinforcement learning, the approach DeepMind used in AlphaGo. Starting from random connections and adjusting them with experience is how deep neural networks are trained, including the language model GPT-2 that OpenAI released in full this month.

## How do Turing's unorganized machines compare with today's neural networks?

They share the key idea: many simple units, connected at first without a plan, that become useful through training rather than through hand-written rules. That is the same starting point a modern network has before it sees any data.

The differences are large, though. Turing's units only ever output 0 or 1, and his connections were either on or off. Modern networks use numbers that can be adjusted by tiny amounts, which lets a method called backpropagation nudge millions of connections at once in the right direction. Turing's networks also had no layers; today's are stacked in many layers, which is where the word "deep" comes from.

The biggest difference is scale and speed. Turing ran his experiments with a pencil and an eraser and wrote that the work was too laborious to push further until electronic machines were available. He hoped a computer could one day run the whole training process while a human looked in occasionally, like a school inspector checking progress. That, in a sense, is what a modern training run is.

## Sources

- Intelligent Machinery: https://weightagnostic.github.io/papers/turing1948.pdf
- Alan Turing: https://plato.stanford.edu/entries/turing/
- Unorganized machine: https://en.wikipedia.org/wiki/Unorganized_machine