Browse Books

Reviewer: H. Van Dyke Parunak

One of the most heated discussions in modern AI concerns the relation of a given artificially intelligent system to the mechanisms that produce similar intelligent behavior in humans. Some researchers focus on the means (so that AI is a mechanism for studying psychological and neurological processes and must seek to mimic those processes), while others focus on the results (so that any mechanism is appropriate so long as the final behavior is “intelligent”). Among those who seek to mimic the process as well as the result, there is a sharp cleavage between the traditionalists, who rely on symbol manipulation in fairly conventional programs and are inspired by psychological studies of human problem-solving behavior, and the connectionists, who concentrate on highly parallel computing architectures with fairly primitive nodes that more or less emulate physiological knowledge of human neural structure and connectivity. Cottrell's 1985 thesis explicitly seeks to imitate human processing and takes a connectionist approach. Hirst's 1983 dissertation [1] has as its primary aim the final performance of the system rather than imitation of human processing, but keeps a close eye on psycholinguistic behavior and follows the traditional symbolic approach. Both studies tackle virtually the same problem, and thus, though they make only limited reference to each other, they offer a tantalizing comparison. Natural language is ambiguous at several levels. Lexically, a single word can correspond to different objects (“ball” is either a fancy dance or a round toy). Syntactically, different parse trees may be possible for the same utterance. For example, in “He saw the man with the telescope,” does “with the telescope” describe the man who was seen or how the observer did the seeing__?__ Semantically, a given portion of an utterance can be mapped in different ways to the semantic objects represented in a parse tree, a problem illustrated by the different roles of “with the telescope” in the two candidate parses of the previous example. The brain differs from classical AI architectures in several ways. For example, it is slower; it tolerates failure of its individual components gracefully; it programs itself rather than requiring programming; and it is massively parallel rather than serial. Connectionism explores the potential of architectures that approximate the brain to produce intelligence. A useful characterization is that connectionist architectures trade time for space. Thus, where Hirst handles semantic ambiguity through a “Polaroid word” that initially records all possible meanings of the word and narrows the set through time, Cottrell defines a separate synthetic neuron for each meaning and lets the network converge until only one neuron, and therefore one meaning, survives. Many connectionists, including Cottrell, argue cogently that the credibility of an artificial system as an explanation for human cognition depends on the similarity of the underlying architectures. The very arguments that Cottrell presents for mimicking human neurophysiology raise important questions for his undertaking. Cottrell's work is not very faithful to brain architecture in two ways. First, he assigns distinct semantic meaning to individual units, though there is strong evidence that the brain represents concepts not as units but as patterns of activation over many units. Second, he engineers the connections to produce the desired behavior, while such connections arise in the course of operation in the brain. Cottrell recognizes these limitations of his work and pleads constraints because of “when and where the work was done.” He argues that the research is still valid as showing “that connectionist models can provide a good basis for cognitive models,” but the motivation for his work seems weak. Each of the three chapters on implementation reviews the problem, surveys available psycholinguistic data on human performance, describes the implementation, and discusses its performance. The research takes the form of a collection of individual experiments, not an integrated system. Each experiment shows the feasibility of a connectionist approach to one fragment of the problem, assuming the other components are available. The more difficult question of whether an integrated solution can be assembled is not addressed experimentally. The implementations draw heavily on human performance data, and Cottrell is candid in his review of their performance against this data. He willingly makes predictions, based on his models, that are testable on human subjects. He covers linguistic studies less thoroughly (Hirst handles this area more completely). For example, the constituent-role grammar on which he builds the syntactic analyzer is well known in linguistic circles as a tagmemic model and is one of several competing formalisms for parsing, but Cottrell does not recognize the broader context of this model or engage the discussion to which it leads. In responding to Etherington and Reiter's criticisms of connectionist networks [2], Cottrell argues that they apply to single-pass marker passing machines like NETL, but not to truly parallel implementations. He represents inference axioms first in default logic and then translates them into bits of network to show the adequacy of a connectionist approach. A critical part of the discussion is tantalizingly incomplete: will a parallel network that implements these designs converge in a correct and consistent fashion__?__ The author <__?__Pub Caret>states that experimental results suggest it will, but he neither describes these results in detail nor offers a formal proof of this vital property. The chapter on aphasia is a nice example of the work's consistent concern with empirical data from human subjects. The chapter on related work has been updated since the original thesis and contains references through 1988. Cottrell's work is a stimulating application of connectionist models to an important problem in natural language, and the existence of Hirst's parallel effort using another methodology offers intriguing tutorial possibilities in reading the two books together. The strongest features of the work are its diligent review of psycholinguistic data and its effort to frame testable predictions in the context of those data. It is unfortunate that Cottrell's specific model has been superseded, and I look forward to future confirmation of its success in more realistic imitations of human neural physiology.