Steven Sloman (1993) Modularity of Mind: a Question Unasked. Psycoloquy: 4(29) Categorization (5)

Volume: 4 (next, prev) Issue: 29 (next, prev) Article: 5 (next prev first) Alternate versions: ASCII Summary
PSYCOLOQUY (ISSN 1055-0143) is sponsored by the American Psychological Association (APA).
Psycoloquy 4(29): Modularity of Mind: a Question Unasked

Book Review of Murre on Categorization

Steven Sloman
Department of Cognitive and Linguistic Sciences
Brown University, Box 1978
Providence, RI 02912


Connectionist models can make substantive contributions to our understanding of the mind in a variety of ways. Murre (1992) provides no evidence that his model makes any of them. He and his colleagues have clearly been busy, but the book is a report of work-in-progress that is lacking in completeness and polish.


Neural networks, neurobiology, psychology engineering, CALM, Categorizing And Learning Module, neurocomputers, catastrophic interference, genetic algorithms.
1. Connectionist models can help us to understand the mind in a variety of ways. Their contribution can be substantive, in providing the means to formulate new scientific theories that explain data that were previously unexplained, in bringing coherence to disparate domains, in providing a unifying theoretical structure, or in providing a more precise account of data than other theories do. Their contribution can also be more subtle. They can provide a new metaphor for thinking about an issue or, relatedly, a new language for theorizing on a topic. Their value in this regard depends on the extent to which they provide an original set of simple, general tools that can be productively combined to generate new theories. Connectionist models are valuable even when their contribution is as vague as that of facilitating communication among different disciplines. Murre (1992) describes a model that purportedly does all these things; I find no evidence that it does any of them.

2. The problem addressed by the title of the book is critical for understanding the architecture of both mind and brain. Why is the brain apparently so modular, and what kind of computational power does that give to processes like learning and categorization? The high point of the book comes in the last chapter when, for two pages, Murre addresses that problem in its general form. There we learn that modularity is important because it makes a model extendible; a new module can be attached without affecting the functioning of the rest of the system. But of course from the point of view of those who believe that processing is wholly interactive, that is just the problem. Every experimental cognitive psychologist has had the frustrating impression at one point or other that everything depends on everything else. Murre shows us some of what you buy by assuming modular structures, such as certain forms of stability, but connectionism is exciting precisely because it gives us a way of describing a system without having to specify functional parts. As others have described in exquisite detail (e.g., Fodor 1983), the extent of modular structure in mental activity is an open theoretical and empirical question.

2.1 This should not bother Murre, because his model is agnostic concerning the structure of our modular architecture. What Murre describes is a particular kind of competitive learning network. This network (or module) can be freely combined with other modules to build a larger network. But when he actually models something, all the effort is put into how these modules are interconnected. Indeed, all the learning occurs in the interconnections. In terms of the theoretical work they perform, the modules, with rare exceptions, play the same role as single units do in standard networks. The modularity structure of his models just adds complexity to his theorizing, making his models seem baroque and confusing -- at least, Murre never presents evidence showing that his model fares better in a direct comparison, in terms of either efficiency or scientific validity, than other simpler models. So as a means of facilitating thinking, theorizing, and communication, these ideas seem to be of little help. The idea of a competitive learning network itself has many virtues, which have been discussed at length by a multitude of theorists, starting with Rosenblatt (1962).

3. The one strong claim that Murre's model does make is that categorization plays a fundamental role in memory. His one argument for the primacy of categorization is that the process of chunking an input into a superordinate category reduces the problem of catastrophic interference. Indeed it does. Of course, it also reduces the advantageous property of automatic generalization given by distributed representations. Moreover, categorization throws away information. Whenever we have fewer superordinate categories than we do inputs (as we must, by the definition of "superordinate"), individuating aspects of inputs become irrelevant if the input is categorized along with other inputs. The extent to which we abstract out commonalities and the extent to which we retain the attributes of individual exemplars is a matter of fierce debate in the literature on categorization, comprehension, and memory. In fact, some of the strongest evidence that we maintain detailed information about superficial features comes from the literature on implicit memory (e.g., Kolers 1975).

4. Murre seems to be confused about what counts as an explanation in science. For example, most scientists agree that a model gains support by being consistent with a variety of data. It gains more support if those data verify a counterintuitive prediction of the model. Murre's model gains little support because it does not make any predictions at all. Both Gluck and Bower (1988) and Kruschke (1992) describe models that do fairly well on both these counts, but Murre dismisses these because the "authors cannot account for most architectural aspects of their model in terms of their psychological function or relation to biological structures" (p. 129).

4.1 Apparently Murre can account for such aspects of his model. To find out, consider the work on memory modeling described in the book. Murre takes as evidence for his model that it implements a theory of explicit and implicit memory tasks developed by Graf and Mandler (1984). They suggested that implicit memory performance is a product of associative learning -- the strengthening of pre-existing associations -- and that explicit memory performance is due to elaborative learning -- the formation of new associations. Although this is an interesting distinction, it has been known for a decade that it is incomplete as a model of the explicit/implicit distinction (see, for example, Schacter 1987). One of several weaknesses is that it fails to explain why implicit memory facilitation sometimes lasts much longer than the aspect of memory measured by explicit tasks. One example can be found in Sloman et al. (1988), who found facilitation on an implicit fragment completion test after 16 months. Murre's failure to consider such evidence is representative of his general approach. He picks and chooses those bits of evidence at any level of analysis and from any domain that happen to be generally consistent with his framework. The few simulations he describes that show some qualitative consistency (and some unmentioned inconsistency) with experimental data in the study of memory is just not sufficient. So much detailed work modeling memory data has now gone on (e.g., Murdock 1974) that it is incumbent on theorists to be qualitatively thorough or quantitatively precise or both when comparing their theories to data.

5. Murre's book reads like a hodge-podge of facts and figures about connectionist models and his particular combination of techniques. He and his colleagues have clearly been busy, but the book is a report of work-in-progress that is lacking in completeness and polish.


Fodor, J. A. (1983). Modularity of mind. Cambridge: MIT Press.

Gluck, M. A. & Bower, G. H. (1988). Evaluating an adaptive network model of human learning. Journal of Memory and Language, 27, 166-195.

Graf, P. & Mandler, G. (1984). Activation makes words more accessible, but not necessarily more retrievable. Journal of Verbal Learning and Verbal Behavior, 23, 553-568.

Kolers, P. A. (1975). Memorial consequences of automatized encoding. Journal of Experimental Psychology: Human Learning and Memory, 1, 689-701.

Kruschke, J. K. (1992). ALCOVE: An exemplar based connectionist model of category learning. Psychological Review, 99, 22-44.

Murre, J.M.J. (1992) Learning and Categorization in Modular Neural Networks. UK: Harvester/Wheatsheaf; US: Erlbaum

Murre, J.M.J. (1992) Precis of: Learning and Categorization in Modular Neural Networks. PSYCOLOQUY 3(68) categorization.1

Murdock, B. B., Jr. (1974). Human memory: Theory and data. Potomac: Erlbaum.

Rosenblatt, F. (1962). Principles of neurodynamics. New York: Spartan.

Schacter, D. L. (1987). Implicit memory: History and current status. Journal of Experimental Psychology: Learning, Memory, and Cognition, 13, 501-518.

Sloman, S. A., Hayman, C. A. G., Ohta, N., Law, J., & Tulving, E. (1988). Forgetting in fragment completion. Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 223-239.

Volume: 4 (next, prev) Issue: 29 (next, prev) Article: 5 (next prev first) Alternate versions: ASCII Summary