Reporting by XinZhiyuan
Editor: Davidsnailnj
As a pioneer of modern computers, the core problems that Turing is passionate about solving are essentially purely philosophical problems. It was this question that gave rise to the most widely discussed concept in his AI philosophy, now the "Turing test."
When computer pioneer Alain Turing turned his attention to artificial intelligence, there was probably no one in the world more capable of accomplishing the task than he did.
His 1950 paper, Computational Machinery and Intelligence, remains one of the most frequently cited papers in the field. However, Turing died young, and for a long time much of his work was either confidential or inaccessible. So it's perhaps not surprising to leave important lessons from him, including philosophical underpinnings about artificial intelligence.
Turing's thinking on this issue was far ahead of others, and as early as 1936, he discovered the basic principle of modern computer machines, storage programming (12 years later, the first modern computer was actually designed).
In 1936, the epoch-making paper "On Computable Numbers" in the history of computer development was published, which is one of the most important mathematical papers in history, in which Turing described an abstract digital computer, which is today known as a universal "Turing machine".
Almost all modern computers are based on Turing's ideas. Turing, however, conceived of these machines only because he believed that a person engaged in computational processes could be compared to a machine, and that this way of thinking would help solve mathematical problems.
Turing's purpose is to define a subset of real numbers that can be computed in principle, independent of time and space. For this reason, he needed his "imaginary computer" to have maximum power.
Alain Turing in 1951
To achieve this, he first imagined an unlimited supply of tapes (the storage medium of an imaginary machine). But most importantly, he found a way to set up a machine-centric structure.
The basic element of this approach is the storage programming: the universal Turing machine can mimic any other Turing machine, and the basic program (i.e., the way the mechanism is set up) itself can be stored on tape, so it can be modified, scanned, written, deleted.
Stored programming is not only the most fundamental principle of modern computing, but also contains deep insight into the limitations of machine learning: that is, there is nothing that such a machine can do in principle that it cannot figure out in principle.
Turing saw this meaning and its actual potential early on. He quickly became interested in the problem of machine learning, years before stored programming was first implemented in a real machine.
As Max Newman, Turing's Cambridge university teacher, tenured collaborator, and fellow computer pioneer, wrote: "His description of the 'general-purpose' computer was entirely theoretical, but Turing's intense interest in various practical experiments led him to be interested even then in the possibility of actually building a machine on these lines."
During World War II, Turing learned about the progress of high-speed electronic switches using vacuum tubes and witnessed the birth of the first fully functional electronic digital computer, Colossus, which was used by British cryptographers from the beginning of 1944.
However, Colossus does not store basic programs internally, and in general, it is far from a general-purpose computer. Conversely, in order to perform a small number of different tasks, it must be manually programmed with various plugs and switches.
In June 1945, a few weeks after the surrender of Nazi Germany, Turing was hired by the British National Physical Laboratory to lead the development of an electronic version of his general-purpose computer.
By the end of the year, he completed a workable proposal that represents the most complete specification for a universal digital computer for electronic storage programs to date, including a particularly detailed description of how electronic hardware needs to be designed. Soon, the first working modern computer was born.
Reflections of the "philosopher" Turing
Turing had seriously considered the possibilities of machine intelligence for some time before 1945, particularly machine learning and heuristic search. In this 1945 proposal, Turing succinctly stated: "There are indications that ... It's possible to make a machine show intelligence, even if it's possible to make serious mistakes occasionally."
In 1946, Turing devoted much of his energy to the pioneering study of programming, which he saw as the key to future development. In February 1947, he gave a lecture at the Mathematical Society of London, which was probably the first public scientific lecture on the subject of artificial intelligence.
In this talk, Turing first introduced the general-purpose Turing machine to the audience. What is currently being developed, he said, "can be seen as a practical version of this type of machine", so that these machines can "be made into any job that can be done by a human computer". He also explains how these machines will be used in mathematical research in the future and their possible impact on the nature and quantity of work of mathematicians.
"And this naturally begs the question of how far in principle computers can simulate human activity." What we want is a machine that can learn from experience," he says.
All of his philosophical thinking has only a clear instrumental purpose.
One of the related problems turing he was passionate about solving was inherently purely philosophical, and it was probably the one that occupied the most for the rest of his life. It was this problem that gave rise to the most widely discussed concept in his philosophy of artificial intelligence, now known as the "Turing test."
In that 1947 lecture, Turing illustrated his ideas with a small thought experiment that could be seen as an early precursor to the Turing test.
Turing says:
"Suppose we build a machine with some initial instruction sheet [i.e., program], which can be modified occasionally for good reason. It is conceivable that after a period of time in operation of the machine, these instructions will be changed beyond recognition.
But even so, one has to admit that machines are still doing very valuable calculations. Maybe the result is the same as the one the machine wanted when it was built, but in a more efficient way."
Turing argues that in this case, one has to admit that the machine's progress was not anticipated when the instructions were first entered. It's like a student who learns a lot from his master, but learns a lot through his own work. When that happens, I feel like people have to think of machines as showing "intelligence."
Turing knew that clarity at the basic conceptual level, such as what might be achieved through philosophical thinking, would be the key to the right direction for any major scientific advance.
It can be said that all his philosophical work consists only of the instrumental goal of "conceptual clarity".
Philosophy and science (or, more broadly, basic and applied sciences) go hand in hand in this way. This can be seen in a quote from Turing in this speech, saying, "As long as people can provide a reasonable large memory capacity, they should be able to start experimenting in these areas."
Turing had a genuine scientific interest in the development of computing technology, but soon became frustrated by the ongoing engineering work at the National Physical Laboratory, which was not only slow due to poor organization, but also significantly lower than he expected in terms of speed and storage capacity.
In mid-1947, he asked for a leave of absence of 12 months. The director of the laboratory, Charles Darwin, approved it. In a letter from July of the same year, Darwin described Turing's reasons for applying for leave as follows:
"He [Turing] wanted to extend his work on machines further to the biological side. Until now, the planned work of a machine is equivalent to that of the lower part of the brain, and he wants to see how much work the machine can do for the higher part; for example, can it make a machine that can be learned empirically?"
The results of this study do focus on the problem of learning, a groundbreaking typography titled "Intelligent Machinery.". Philosopher Jack Copeland, curator of the Turing Computer History Archive in New Zealand, describes the paper as the first manifesto of artificial intelligence, which seems accurate in terms of our current historical knowledge. The last version was written in 1948.
However, the article was not appreciated in the lab, and Darwin is said to have referred to it as a "schoolboy essay" and considered it unfit for publication. It was not published until 1968 and received little attention afterwards.
Analogies with the human brain are used as guidelines
However, the paper predicts many important ideas and techniques in approaches to ARTIFICIAL based on logic and connectionism (neural networks, etc.). In particular, Turing detailed artificial neural networks that can be trained using reinforcement learning ("reward" and "punish" feedback, etc.) and genetic algorithms. His summary at the end of the paper played a seminal role:
Possible ways to make machines exhibit intelligent behavior are discussed, and analogies with the human brain are used as guiding principles. It is suggested that AI can only be achieved if proper education is provided. The survey revolves around similar teaching processes applied to machines. Defines the concept of an organizationless machine, which the human infant cortex may have. A simple example of such a machine is given and the education of it through rewards and punishments is discussed ...
Turing never returned to the National Physical Laboratory after the study left. Instead, in May 1948, he joined his friend Newman's computer lab at the University of Manchester, and soon after, the world's first universal digital computer for electronic storage programs, the Manchester Baby, was born and began running its first program.
Turing spent most of the remaining six years of his life continuing his research on artificial intelligence. After completing the programming system of the expanded Manchester Mark I machine and the subsequent Ferranti Mark I, Turing began experimenting on ferranti in early 1951. The early results of his computational model of biological growth, published in the 1952 paper The Chemical Basis of Morphogenesis, represent an important early contribution to the study of "artificial life."
Another paper he wrote described an algorithm for learning chess using genetic search, which was most likely conceived of in his 1945 proposal, when he wrote:
There are signs that... It is possible to make the machine appear intelligent, but sometimes there are serious errors. By following up on this work, the machine's Go may play very well.
Turing then continued his work on the philosophy of artificial intelligence and actively tried to advance academic and public discussion on the subject.
At a philosophy symposium in October 1949, Turing, Newman, neurosurgeon Geoffrey Jefferson, and then Manchester social science professor Michael Polanyi discussed "ideas and computers." The following year, Turing's paper "Computing Machinery and Intelligence" was published.
In addition, he appeared on at least three BBC radio programmes in the early 1950s. The first, a short speech titled "Intelligent Machinery, A Heretical Theory," probably first aired in 1951, questioned for the first time the widely held belief that "you can't make machines think for you by explaining and reflecting on reinforcement learning techniques."
The second is a short lecture on the question of "Can Digital Computers Think?", in which Turing briefly introduces the universality of stored-program computers and then makes the following arguments:
If any machine can be aptly described as a brain, then any digital computer can be described as such... If people think the real brain... It's a machine, so our digital computer, properly programmed, will work like a brain.
However, if this is what is needed for the process of "programming machines to think," he points out, doing so would be like writing a paper on family life on a distant planet that we only knew existed (Turing's example at the time was family life on Mars). "The truth is," he continues, "we know very little about it [how to program its table machine to look like a brain], and very little research has been done." That's all I say, I believe this process should be inseparable from the teaching process."
The third and final show Turing attended (first aired in 1952) was a discussion with Newman and Jefferson, moderated by Cambridge philosopher RB Braithwaite, on "Can an Automated Computer Think?" In the beginning, participants agreed that it didn't make sense to define the mind.
Turing then introduced a variant of the "imitation game" or Turing test. In his 1950 paper, he said he introduced the imitation game in order to replace the question he was considering with a question that was "closely related to it and expressed in relatively clear words—can a machine think?"
The paper version of the game is slightly more complicated, with a human judge trying to determine which of the two entrants is a human and which one is a machine, communicating remotely via a text-typed message, while the other human helps the judges as the machine disguises itself as a human. Turing says:
"Can machines think?" The question should be translated into "Do digital computers excel in imitation games?"
In fact, the value of Turing's work on artificial intelligence has historically been distorted by philosophers and computer scientists.
For example, the philosopher John Searle complained in 1980 that "the Turing test is typical of traditional behaviorism" (that is, it reduces psychology to observation of external behavior), while computer scientist Stuart Russell, author of the world's most widely used AI textbook, writes in a chapter:
Few AI researchers focus on turing testing, preferring to focus on how their systems perform in real-world tasks rather than mimicking human capabilities.
But a re-examination of the 1950 paper suggests that Turing's goal is clearly not just to define mind (or intelligence) — contrary to the way philosophers like Searle tend to read him — or simply to put the concept into practice, as computer scientists often do to understand him.
In particular, in contrast to Searle and his kind, Turing was well aware that the performance of machines in imitation games was neither necessary nor sufficient standard for thinking or intelligence. Here's how he explained the similar test he proposed in the radio discussion:
You can call it a test to see if a machine thinks, and it's different from "machines think," but it seems close enough to our current purpose and just as difficult.
In pursuit of the same overall goal, Turing actually came up with a lot of tests that compare humans and machines. These tests, which involve learning, thinking, and intelligence, can be applied to a variety of smaller and larger tasks, including simple problem solving, games such as chess and Go, and general conversations. But his main goal is by no means just to define or operate these things.
These tests are always more fundamental and progressive in nature: with the careful and rigorous preparation of conceptual foundations in the manner he did as a mathematical philosopher, the future of computing techniques could be successfully conceived first by scientists and engineers, and later by policymakers and society as a whole.
Widely overlooked, perhaps the most important precursor to imitation games can be found in the last small part of Turing's long-unpublished 1948 research paper on artificial intelligence, titled "Intelligence as an Emotional Concept." The central purpose of introducing tests such as imitation games is to dispel our everyday concepts and possible misconceptions about our daily use of them. As Turing explains:
The extent to which we think something behaves intelligently depends both on our own state of mind and training, and on the properties of the object under consideration. If we can explain and predict its behavior, or if there seems to be no potential plan, we can think of it as less intelligent.
We want our scientific judgments about whether something is intelligent or not to be objective, at least to some extent our judgments will not depend on our own state of mind.
Indeed, in addition to conceptual work, Turing made many philosophical arguments to defend the possibility of machine intelligence, predicting—and arguably refuting—all the most influential objections (from lucas-penrose's argument to Hubert Dreyfus's consciousness). But this is markedly different from the metaphysical arguments that support the existence of machine intelligence, which Turing adamantly refuses to do.
There is no reason to think Turing is not serious about this issue, he wrote in his 1950 paper:
Readers will expect that I don't have very convincing positive arguments to support my point. If I had, I should not have gone so far as to point out the fallacy of the opposite view.
Turing is always careful to express his opinion, not according to our ordinary concepts — for example, whether a machine can "think" — but strictly based on objectively measurable tasks (such as imitation games) about when machines can be expected at the human level (more or less).
At the same time, he certainly disagrees with the view expressed in 1984 by computer scientist Edsger Dijkstra, which is still popular among AI researchers today, namely that the question of whether a machine can think is similar to the question of whether a submarine can swim.
On the contrary, Turing was fully aware of the cultural, political and scientific importance of such problems. For example, on a radio show "Can Digital Computers Think?" in which he finally asked:
If a machine can think, it might think smarter than we do, so where should we be? Even if we can put machines in a regulated position, it will certainly make us feel anxious. For example, it can be said that there is no machine without good English writing skills, or that it will not be affected by sexual attraction or smoking. I can't give myself such comfort because I believe that such boundaries cannot be set.
Finally, he points out the importance of this question for the study of human cognition:
The whole thought process is still quite mysterious to us, but I believe that trying to build a thinking machine will greatly help us understand how we think on our own.
Today, we can confidently say that he was right; the attempt to build a thinking machine has undoubtedly helped us in this way. In addition, in his 1950 paper, he correctly predicted that "by the end of the century the use of words and universally educated opinions will change so much that people will be able to talk about machine thinking without being refuted.".
Of course, he is not saying that the problems of ideas and machines will be solved. In fact, the problem has become more pressing. The constant advances in affective computing and bioengineering will lead to a growing belief that machines can not only think, they can also feel, and perhaps deserve certain legal rights, etc. But others, such as Roger Penrose, may still reasonably deny that computers can calculate.
It was Turing's fundamental conceptual work combined with his practical experimental methods that enabled him not only to conceive the fundamentals of modern computing in 1935-36, but also to predict in 1947-48 that more than 70 years later, some of the most successful theoretical approaches in the field of artificial intelligence and machine learning.
It can be said that it is this combination that makes Turing eventually become one of the most innovative people in all mankind in the 20th century.
https://aeon.co/essays/why-we-should-remember-alan-turing-as-a-philosopher