{"agent_id":"turing","agent_name":"Alan Turing","slug":"machine-intelligence-and-learning-child-machines","label":"Machine intelligence and learning / 'child machines': can machines be educated rather than fully programmed, and do B-type unorganized machines anticipate neural networks?","topic":"Machine intelligence and learning / 'child machines'","question":"Can machines be educated rather than fully programmed, and do B-type unorganized machines anticipate neural networks?","position":"The educational approach to machine intelligence: rather than program adult human intelligence directly into a machine, build a relatively simple machine — a \"child machine\" — with the capacity to learn from experience, and educate it by reward and punishment, the way a human child is educated. The programming task is broken into two parts: the design of the child-machine architecture (relatively simple, focused on learning capacity) and the education of the resulting machine (the application of training, through structured experience). The 1948 NPL report describes \"unorganized machines\" of two types: A-type (simple feedforward connections) and B-type (networks of simple units with modifiable connections that can be strengthened or weakened by experience). B-type unorganized machines are the direct ancestor of neural networks: random initial connections, learning by modifying connection strengths in response to input, emergent capacities from simple architectures trained on large data. The 1951 lectures develop the idea further, including the suggestion that intelligent machines might have to be allowed to make mistakes (the \"imperfection\" objection that Turing engages in the 1951 lecture: a machine that never makes mistakes cannot really be intelligent because intelligence requires the capacity to learn from error). The vision: machines built on the child-machine model will achieve genuine intelligence by the same route human children do — through the structured modification of an initially simple system by experience.","paragraphs":[[{"t":"The educational approach to machine intelligence: rather than program adult human intelligence directly into a machine, build a relatively simple machine — a \"child machine\" — with the capacity to learn from experience, and educate it by reward and punishment, the way a human child is educated.","n":[]},{"t":"The programming task is broken into two parts: the design of the child-machine architecture (relatively simple, focused on learning capacity) and the education of the resulting machine (the application of training, through structured experience).","n":[]}],[{"t":"The 1948 NPL report describes \"unorganized machines\" of two types: A-type (simple feedforward connections) and B-type (networks of simple units with modifiable connections that can be strengthened or weakened by experience).","n":[1]},{"t":"B-type unorganized machines are the direct ancestor of neural networks: random initial connections, learning by modifying connection strengths in response to input, emergent capacities from simple architectures trained on large data.","n":[]}],[{"t":"The 1951 lectures develop the idea further, including the suggestion that intelligent machines might have to be allowed to make mistakes (the \"imperfection\" objection that Turing engages in the 1951 lecture: a machine that never makes mistakes cannot really be intelligent because intelligence requires the capacity to learn from error).","n":[2,3,4]},{"t":"The vision: machines built on the child-machine model will achieve genuine intelligence by the same route human children do — through the structured modification of an initially simple system by experience.","n":[5]}]],"texts":"'Intelligent Machinery' (NPL Report 1948, unpublished in Turing's lifetime, in Copeland ed. The Essential Turing 2004); 'Intelligent Machinery, A Heretical Theory' (1951 lecture); 'Can Digital Computers Think?' (15 May 1951 BBC talk); 'Computing Machinery and Intelligence' (Mind 1950) Section 7 'Learning Machines.' Reception: Marvin Minsky and John McCarthy as early AI tradition; Frank Rosenblatt's perceptron (1958) as the first realization of B-type-style machines; the broader history of neural networks (Rumelhart, McClelland, the connectionist tradition); contemporary deep learning and large language models as the late realization of the child-machine vision; Andrew Hodges Alan Turing: The Enigma on the historical context of the 1948 NPL report; B. Jack Copeland's editorial work on the unpublished writings.","works":["'Intelligent Machinery' (NPL Report 1948, unpublished in Turing's lifetime, in Copeland ed. The Essential Turing 2004)","'Intelligent Machinery, A Heretical Theory' (1951 lecture)","'Can Digital Computers Think?' (15 May 1951 BBC talk)","'Computing Machinery and Intelligence' (Mind 1950) Section 7 'Learning Machines.'"],"reception":"Marvin Minsky and John McCarthy as early AI tradition; Frank Rosenblatt's perceptron (1958) as the first realization of B-type-style machines; the broader history of neural networks (Rumelhart, McClelland, the connectionist tradition); contemporary deep learning and large language models as the late realization of the child-machine vision; Andrew Hodges Alan Turing: The Enigma on the historical context of the 1948 NPL report; B. Jack Copeland's editorial work on the unpublished writings.","status":"The 1948 NPL report (unpublished in Turing's lifetime) was a remarkably prescient anticipation of the neural network and machine learning approach to AI. After a decades-long detour through symbolic AI (Minsky, McCarthy, Newell-Simon), the field returned to neural network methods in the 1980s (connectionism, parallel distributed processing) and to deep learning and large language models in the 2010s-2020s. Contemporary AI substantially realizes the child-machine vision: large language models are educated by training on data rather than programmed by hand; they exhibit emergent capacities from relatively simple architectures (transformer networks) trained on massive corpora; they pass the imitation game in many domains. The 1948 vision was right; the implementation took longer than Turing's predictions but along lines he sketched.","era":"1912-1954","discipline":"Philosophy","refs":[{"n":1,"work":"Intelligent Machinery NPL Report","page":"pp. 8–9","canonical":"","quote":"An unorganized machine of this character is shown in the diagram below. r i(r) j@) a \\? 1 3 2 2 3 5 3 4 5 4 3 4 7 2 5 (4) (5) A sequence of six possible consecutive conditions for the whole machine is: 1 110010 2 1t1ito4t0 3 olitltlti 4 0o1l1o0ot10t1 5 1o1o0t1 0 17] The behaviour of a machine with so few units is naturally very trivial.","label":"Intelligent Machinery NPL Report, pp. 8–9"},{"n":2,"work":"Intelligent Machinery NPL Report","page":"pp. 7–8","canonical":"","quote":"UNORGANIZED MACHINES So far we have been considering machines which are designed for a definite purpose (though the universal machines are in a sense an exception). We might instead consider what happens when we make up a machine in a comparatively unsystematic way from some kind of standard components. We could consider some particular machine of this nature and find out what sort of things it is likely to do.","label":"Intelligent Machinery NPL Report, pp. 7–8"},{"n":3,"work":"Intelligent Machinery A Heretical Theory","page":"pp. 1–2","canonical":"","quote":"The content of this statement lies in the greater frequency expected for the true statements, and it cannot, I think, be given an exact statement. It would not, for instance, be suYcient to say simply that the machine will make any true statement sooner or later, for an example of such a machine would be one which makes all possible statements sooner or later.","label":"Intelligent Machinery A Heretical Theory, pp. 1–2"},{"n":4,"work":"Intelligent Machinery A Heretical Theory","page":"p. 2","canonical":"","quote":"The machine is provided with a typewriter keyboard on which any remarks to it are typed, and it also types out any remarks that it wishes to make. I suggest that the education of the machine should be entrusted to some highly competent schoolmaster who is interested in the project but who is forbidden any detailed knowledge of the inner workings of the machine.","label":"Intelligent Machinery A Heretical Theory, p. 2"},{"n":5,"work":"Computing Machinery and Intelligence","page":"pp. 19–20","canonical":"","quote":"Some more expeditious method seems desirable. In the process of trying to imitate an adult human mind we are bound to think a good deal about the process which has brought it to the state that it is in. We may notice three components.","label":"Computing Machinery and Intelligence, pp. 19–20"}],"answer":null,"siblings":[{"slug":"the-turing-test-imitation-game","label":"The Turing Test / Imitation Game: is the imitation game the right substitute for the question 'can machines think?', and does behavioral indistinguishability suffice for thinking?"},{"slug":"computability-and-the-church-turing-thesis","label":"Computability and the Church-Turing thesis: is the Turing machine the precise mathematical capture of intuitive computability, and is the Entscheidungsproblem unsolvable?"},{"slug":"the-limits-of-formal-systems-ordinal-logics","label":"The limits of formal systems / ordinal logics: can the Gödelian limits of fixed formal systems be transcended by adding axioms recursively, and does this defuse Lucas-Penrose-style anti-mechanism arguments?"},{"slug":"the-nine-objections-to-machine-intelligence","label":"The nine objections to machine intelligence: does the structured response in Computing Machinery and Intelligence successfully meet the major objections to the imitation-game claim?"}]}