The Role Of Sensorimotor Function Associative Memory And Reinforcement Learning In Automatic Acquisition Of Spoken Language By An Autonomous Robot
The Role of Sensorimotor Function, Associative Memory and Reinforcement Learning in Automatic Acquisition of Spoken Language by an Autonomous Robot
Stephen E. Levinson
The Strong Theory of AI is clearly expressed in Turing's seminal 1950 paper in which he proposes the infamous ``Turing Test'' for intelligent behavior. The intuition behind his elaborate argument is that the Universal Turing Machine is capable of performing virtually any symbol manipulation process and is therefore sufficient for creating a mental model of the world. The preferred realization of this idea is a synthetic model, complete in every detail, that computes a symbolic representation of meaning from natural language text. This process is to be based on predetermined operations on predefined symbols. The symbols and operations are to be determined by some combination of scientific observation, introspection and divine inspiration.
In the penultimate paragraph of this paper, Turing offers an astounding and often overlooked alternative suggesting that the symbols and relations amongst them could be learned from real-world sensory data. In fact, he urges that both approaches be tried. It may be argued that both approaches conform to the Strong Theory of AI but in significantly different ways. The direct synthesis is predicated on a discrete symbolic model perfectly isomorphic to reality and unaffected by any uncertainty present in the physical world. This approach assumes that the sensorimotor periphery may safely be ignored. The alternative theory uses the computational power of the Turing Machine to analyze the physical processes from which distributed symbols and structures, derive. Thus, cognitive function emerges from physical measurement and mathematical description of the experience of and participation in reality.
A unification of these two complementary interpretations of the Strong Theory leads to the following hypothesis about brain, mind and language. The disembodied mind is a fantasy. Thought is almost exclusively the product of the vast associative memory called the brain. The memory is able to capture spatiotemporal order and represent it episodically. Thus there can be no isolated perceptual or cognitive functions. Memory is built up from instincts by the reinforcement of successful behavior in the real world at large. As a cognitive model of reality is acquired, a linguistic image of it is formed primarily in response to semantic information. Other levels of linguistic structure exist to make semantics robust to ambiguity. When the language is fully acquired, most mental processes are mediated linguistically and we appear to think in our native language which we hear as our mind's voice.
We are now testing this hypothesis experimentally by constructing an autonomous robot and training it by reinforcement methods to ascertain to what extent it is capable of acquiring language. To date the robot has attained some simple visual navigation and object manipulation abilities which it can perform under spoken command.