Successful cognitive architectures possess subjective and intersubjective meaning: They make cognition comprehensible to individual cognitive scientists, and they organize like-minded cognitive scientists into genuine communities. These meaning dimensions are not emphasized in conventional analyses of architectures (Anderson, 1990; Marr, 1982; Newell, 1990; Pylyshyn, 1984; Rumelhart & McClelland, 1986). However, they are consistent with philosophies of science that portray graduate training as enculturating budding scientists into scientific communities (Kuhn, 1996). Through this training, scientists acquire the paradigmatic principles of a theoretical worldview (Goodman, 1968; Hanson, 1958) that provides a “group-licensed way of seeing” (Kuhn, 1996, p. 189). Architectural worldviews structure the computational, theoretical, and empirical research activities of cognitive scientists.
3.1. Architectural worldviews
Understanding the subjective meaning of a cognitive architecture requires going beyond reading scientific papers. It requires internalizing the architecture’s worldview—the distinctive information processing style it attributes to cognition. Tutorials must be worked through, modeling exercises completed, and simulation environments mastered. This can require a year or more of sustained effort, typically spent during graduate school or during a post-doctoral or sabbatical year. Through these efforts, cognitive scientists learn to see cognition through the architecture’s worldview (Wittgenstein, 1958, pp. 194–197) and to design computational models that embody its characteristic information processing style.
Understanding the intersubjective meaning of a cognitive architecture means understanding the research activities of other cognitive scientists who share the same architectural worldview. By apprenticing in an architectural community, cognitive scientists learn from conversations with other members of the community. They inspect and adapt computational models written by others, and they share their own models and modify them based on feedback provided by more experienced members of the community (Newell, 1990, pp. 504–505). In this way, an architecture becomes “a language spoken among all members of the community, rather than a language spoken by authors of the theory to readers of the theory” (Anderson, 2007, p. 41).
The development of subjective and intersubjective meaning is illustrated by the growth of the Soar architecture and community. Through the mid-1980s, Soar was a project local to Carnegie Mellon University, to the tight research group that formed around Newell, Laird, and Rosenbloom. Soar 4 was released in 1986 with the explicit goal of attracting a larger architectural community (Laird & Rosenbloom, 1996, p. 31). The architecture was ported to Common Lisp so that it could run on a range of workstations, the user interface was improved, and a manual was written. These changes made it possible for cognitive scientists at other institutions to use Soar, and to develop a subjective understanding of it. To foster intersubjective understanding of Soar, regular workshops were organized. (As of 2010, 30 have been held.) Soar developers and users also formed a virtual community, knitted together by email and other communication technologies (Carley & Wendt, 1988). The growth and maintenance of the Soar community was important enough that Newell was willing to pay large “managerial costs,” spending “over half of every day on it” (Newell quoted in Agre, 1993, p. 442).
An expansion of subjective and intersubjective meaning also accompanied the re-emergence of connectionist architectures in the 1980s. The availability of the Parallel Distributed Processing tutorial and simulation software (McClelland & Rumelhart, 1988) enabled cognitive scientists working alone at their personal computers to move beyond reading about the models of others and to construct their own models, developing their subjective understanding of connectionist architectures. The founders of the connectionist revolution were cognizant of the need to grow to an architectural community. They organized conferences such as Neural Information Processing Systems (first held in 1987) and workshops such as the Connectionist Models Summer School (first held in 1986). This enabled cognitive scientists outside the University of California—San Diego and other hotspots “to make contact with their colleagues,” and to develop an intersubjective understanding of connectionist architectures (Mozer, Smolensky, Touretzky, Elman, & Weigand, 1994).
Finally, the success of ACT-R has been characterized by the growth of subjective and intersubjective meaning (Anderson, 2007, p. 41). The monograph that introduced ACT-R (Anderson, 1993) included a Common Lisp implementation of the architecture, a reference manual, and a tutorial for beginners. The software has been ported to a number of platforms over the years and the documentation refined. The widespread availability of these materials enables cognitive scientists working alone at their computers to develop a subjective understanding of ACT-R. A number of institutional structures have also been established, and around them a genuine ACT-R community has grown. An annual summer school trains new members in proper usage of the architecture, and an annual workshop speeds dissemination of domain models throughout the community. The community gains additional coherence through internet-based technologies: A website serves as a repository for its documentation, papers, and models; a mailing list enables newcomers to query experienced users about problems that arise during model construction; and so on. These pedagogical materials and institutional structures make it possible for cognitive scientists to develop a deeper subjective and intersubjective understanding of ACT-R than was possible of its predecessors.
The subjective and intersubjective meaning of cognitive architectures contribute to the worldviews they offer on human information processing. These worldviews are selective, magnifying some aspects of cognition for closer inspection while shrinking others into the infinite distance. With this selectivity comes generativity: fresh perspectives on cognitive phenomena, and their explanation as computational phenomena. Generative architectures enable cognitive scientists to glimpse “the subtlest and most esoteric of the phenomena” (Kuhn, 1996, p. 164). The generativity of architectures has been discussed by other cognitive scientists (Newell, 1990, p. 14; Pylyshyn, 1984, p. 128). For example, Anderson (1983) downplays conventional evaluative criteria such as the empirical coverage an architecture offers and instead emphasizes “the success, or fruitfulness, of the theories it generates” (p. 12).
Generativity takes two forms. Prospectively generative architectures open up new ground for exploration. For example, consider the rise of symbolic architectures during the 1960s and 1970s. These architectures offered fresh perspectives on high-level forms of cognition, such as the notion of problem solving as heuristic search through problem spaces (Newell & Simon, 1972). The symbolic worldview also supplied empirical methods for documenting new phenomena. One example is protocol analysis, which enables researchers to track how participants explore problem spaces as they solve problems, comprehend discourse, and so on (Newell & Simon, 1972; Pressley & Afflerbach, 1995). The result was a spate of new models, theories, and empirical regularities (Ericsson & Simon, 1993).
Retrospectively generative architectures offer new perspectives on familiar terrain. Looking at well-known empirical regularities through a new architectural worldview often suggests novel theoretical and computational accounts. Consider the explosion of connectionist architectures during the 1980s. A number of unexplained empirical regularities had accumulated during the first part of this decade, for example, from the emerging literature on the cognitive impairments of neuropsychological patients (Caplan, Baker, & Dehaut, 1985; Cohen & Squire, 1980; Warrington & Shallice, 1984). The symbolic architectures that dominated cognitive science at the time failed to offer insightful accounts of these findings. In this vacuum, the generativity of the connectionist worldview was breathtaking. For a time, it seemed as if skilled connectionists needed only glance at a domain, such as the cognitive impairments of people with dyslexia (Plaut & Shallice, 1993) or schizophrenia (Cohen & Servan-Schreiber, 1992), to generate fresh computational insights.