Christopher D. Green (1998) Does Localism Solve Connectionism's Problem?. Psycoloquy: 9(14) Connectionist Explanation (11)

Volume: 9 (next, prev) Issue: 14 (next, prev) Article: 11 (next prev first) Alternate versions: ASCII Summary
PSYCOLOQUY (ISSN 1055-0143) is sponsored by the American Psychological Association (APA).
Psycoloquy 9(14): Does Localism Solve Connectionism's Problem?

Reply to Grainger & Jacobs on Connectionist-Explanation

Christopher D. Green
Department of Psychology
York University
Toronto, Ontario M3J 1P3


Grainger & Jacobs (1998) argue that the problems facing connectionism discussed in my target article (Green 1998) can be overcome by switching to a localist connectionist perspective. I question whether the cost of doing so outweighs the disadvantages of staying with the parallel distributed processing approach to connectionist cognitive science.


artificial intelligence, cognition, computer modelling, connectionism, epistemology, explanation, methodology, neural nets, philosophy of science, theory.
1. Grainger & Jacobs (1998) suggest that localist connectionism (LC) avoid the problems faced by parallel distributed processing (PDP), and identified in my target article (Green 1998), because it is interpretationally transparent; i.e., it is clear what each element in the theory refers to in the real world. On the surface, this appears to be a great advantage of LC over PDP, but I have some concerns about the costs involved.

2. First, I plead guilty to Grainger & Jacobson's charge of having used the term "connectionism" to refer to PDP models. I, of course, understand that "connectionism" covers a range of architectures that extend well beyond the boundaries of PDP. I must confess, however, that I was not aware that serious work was still being conducted on LC. It conjures up images of the semantic networks popularized by Collins and Quillian (1969) which, conceptually valuable as they may once have been, no longer earn the allegiance they once did.

3. That having been confessed, it must be clear that I am not aware of the details of the work presented in Grainger & Jacobs' new book (1998b), I cannot give much by way of informed reply. I do wonder about the a few issues, however. PDP has provided such a strong challenge to traditional symbolic cognitive science, inter alia, because it is easily able to simulate features of real behavior with which symbolic models have significant trouble: (1) learning a new domain from scratch (i.e., without significant pre-programmed (i.e., "innate") information about the domain), (2) generalizing what is learned to new inputs never before processed, (3) degrading "gracefully" when faced with incomplete input or internal damage.

4. All these advantages accrue to PDP as a direct result of the distributed nature of their representations (or so it has long been thought). LC models, in contrast, not having distributed representations, have not traditionally had these advantages, and have hence slipped from the limelight in favor of PDP models. If Grainger and Jacobs have overcome these disadvantages of LC, then that must be considered to be a great breakthrough (and my guess would be that this could then be easily transfered to traditional symbolic models as well, about which I do not share Grainger & Jacobs's qualitative-quantitative qualms).

5. If not, however, then the price LC models have paid for their interpretational transparency is very great, and I doubt that many PDP connectionists will be returning to them to model cognition.


Collins, A.M. & Quillian, M.R. (1969) Retrieval time from semantic memory. Journal of Verbal Learning & Verbal Behavior, 8, 240-247.

Grainger, J. & Jacobs, A. M. (1998a) Localist connectionism fits the bill: Commentary on Green on Connectionist-Explanation. PSYCOLOQUY 9(9)

Grainger, J & Jacobs, A. M. (Eds.) (1998) Localist Connectionist Approaches to Human Cognition. Mahwah, NJ.: Erlbaum.

Green, C. D. (1998) Are connectionist models theories of cognition? PSYCOLOQUY 9(4)

Volume: 9 (next, prev) Issue: 14 (next, prev) Article: 11 (next prev first) Alternate versions: ASCII Summary