In his paper on “Computing Machinery and Intelligence”, Alan Turing raised two interesting research challenges. The first one, “The Imitation Game” commonly called “The Turing Test” has received substantial notoriety but has not been subjected to careful research experimentation. The second problem, that of creating a Child Machine and educating it, has been largely ignored, possibly because it got overshadowed by the hoopla surrounding the Turing Test. I believe both these topics continue to be worthy of research attention. This talk attempts to broaden Turing’s proposals from a perspective of half a century of advances in AI, CS and Brain Science.
The original imitation game proposed by Turing attempts to resolve the question of whether a human or a machine is at the other end of the Teletype cable. While this provides for an ingenious validation, it is too narrow a view to truly reflect the full complexity of the nature of human intelligence. We now know that intelligence is more than answering questions. It is difficult to say that people in a society of 5000 years ago that invented tools and writing are any less intelligent just because they could not read and write or answer typed questions. A modern version of Turing’s intent is provided by Jim Gray in his Turing Award Talk. He proposes several attributes for human level intelligence such as see, hear and speak as well as a native speaker.
Note that even an illiterate person in Kalahari Desert can perform these tasks. In spite of 50 years of research, getting a computer to see, hear and speak has proved to be difficult. However there are glimmers of hope from advances of the last decade on problems previously thought to be decades away from solution. These include Google’s language translation systems, Apple’s Siri speech understanding system providing answers to spoken questions, and IBM’s Watson answering questions on Jeopardy. All these advances seem to point to brute-force approach to Intelligence rather than some concise and elegant paradigm.
Surprisingly, there has been little progress in pursuing the research implied by Turing’s second suggestion of realizing AI by creating a Child Machine and educating it.
McCarthy in his paper “The Well Designed Child” and Minsky in his books “The Society of Mind” and “The Emotion Machine” make several important suggestions on the attributes and properties of such a machine, but not at the level of what a computer architect might do.
Given 100 million computational elements and 100 million wires to connect them, what would a Processor-Memory-Switch architecture diagram look like? How would such a system self-adopt and self-program itself to recognize and react to various sensory inputs? How would it accept and respond to an external teacher, and dynamically modify its memory structures? Turing provides possible approaches to these problems given the technologies of 1950s. In the ensuing 50 years, no one seems to have considered Turing’s proposals as serious, Ray Kurzweil notwithstanding.
The task of deciding on the architecture of a Child Machine, which is able to learn from sensory motor experiences and self-organize and evolve into adult intelligence, is daunting. Some necessary conditions are:
Recent advances in Brain Science yield many insights into the innate architecture of a baby brain at birth.
Educating a Child Machine presumes that it has built-in mechanisms for learning architecture. It should enable indexing and organizing sensory-motor data continuously so as to store, match, recognize, retrieve, and change internal memory structures dynamically based on the signals. We begin with the following assumptions
Once educatable, the Child Machine would be exposed to several thousand hours of raw speech and language data including clarification and correction by a teacher. The primary objective is to endow a Child Machine with speech and language understanding and spoken language output of a 2 to 4 year old child, recognizing and generating full sentences with a vocabulary of around 1,000 words and phrases. Usually, most proposals do not consider survival and emotion to be part of Intelligence at birth, but should be included because goal directed emotional behavior such as joy, grief, fear, envy, greed, anger, and instinct for survival may be essential part of being intelligent.
For now we can assume that we are not constrained by biological constraints in choosing the hardware for Child Machine; it can be assumed we have unlimited memory, computation and bandwidth. If Newell and Simon were around here today they might wish to the hardware for Child Machine. To them it was equally important to use the design and education of a Child Machine, as an Information Processing Model of a Human Child and to create a Constructive Cognitive Theory of how the Human Brain works. For both of them, AI WAS “theoretical psychology”. To each his own. My point is, instead of trying to “produce a programme to simulate the adult mind” with intellectual capabilities of PhDs, we should embrace Turing’s suggestion of first try to create a Child Machine that has the abilities of every man, woman and child on the planet, literate or not. If we can create a Child Machine which can see, hear and speak as well as a 4 year old by educating it, we can probably educate it to learn everything else too! We can only see a short distance ahead, but we can see plenty there that needs to be done. Thank you.