Jonathan Grainger (1998) Localist Connectionism Fits the Bill. Psycoloquy: 9(10) Connectionist Explanation (7)

Volume: 9 (next, prev) Issue: 10 (next, prev) Article: 7 (next prev first) Alternate versions: ASCII Summary
PSYCOLOQUY (ISSN 1055-0143) is sponsored by the American Psychological Association (APA).
Psycoloquy 9(10): Localist Connectionism Fits the Bill

Commentary on Green on Connectionist-Explanation

Jonathan Grainger
Centre de Recherche en Psychologie Cognitive, CNRS
Universite de Provence

Arthur M. Jacobs
Dept. of Psychology
Philips University of Marburg,
Marbug, Germany


Green (1998) restates a now standard critique of connectionist models: they have poor explanatory value as a result of their opaque functioning. However, this problem only arises in connectionist models that use distributed hidden unit representations, and is NOT a feature of localist connectionism. Indeed, Green's critique reads as an appeal for the development of localist connectionist models as an excellent starting point for building a unified theory of human cognition.


artificial intelligence, cognition, computer modelling, connectionism, epistemology, explanation, methodology, neural nets, philosophy of science, theory.
1. First, if we agree that theory development in psychological science is ready for the shift from prequantitative verbal-boxological modeling toward more formal modeling efforts, then the kinds of questions we should be asking are: What kind of quantitative modeling is appropriate? How should we evaluate its appropriateness? In other words, the verbal theories of human memory discussed by Green (1998) are not a serious alternative to whatever connectionism might offer. They are at best a starting point for developing more formal accounts of human memory. We have recently argued that localist connectionism provides a promising framework for such an endeavor (Grainger & Jacobs, 1998).

2. Green (1998), as well as many other critics of connectionism, appears to use the term connectionism as synonymous with trainable networks with hidden units (often called PDP models, and typically trained with backpropagation, Rumelhart, Hinton, & Williams, 1986). Many connectionist models do not include hidden units. Some of these are trainable (with Hebbian learning, for example), and some are hardwired (e.g., McClelland & Rumelhart's, 1981, interactive activation model). We refer to any connectionist model in which all processing units can be unambiguously assigned a meaningful interpretation as "localist connectionist." Note that, as in all connectionist models, all processing units in localist connectionist models are identical; it is only their position in the network that guarantees their unique interpretation. The modeler can artificially label each of these units in order to facilitate interpretation of network activity.

3. Grainger and Jacobs (1998) analyzed the advantages of adopting a localist connectionist approach as opposed to the currently more popular PDP approach. Here we will discuss only those points relevant to the issues raised by Green (1998). Green identifies the close connection between theoretical and observable entities as a critical feature of traditional scientific theories. One must be able to link transparently the theoretical entities of the theory to the observable entities in the target world in order to achieve explanatory adequacy. Without examining the extent to which this is fails to be a feature of PDP models, it should be clear from the above discussion that localist connectionist models do provide this transparent link. Units in localist connectionist models do refer to relatively uncontroversial aspects of the target world. They represent the categories (such as letters and words) that the brain has learned from repeated exposure to the environment.

4. As noted by Jacobs, Rey, Ziegler, and Grainger (1998), transparency will always tend to diminish as models become more complex. Jacobs et al. conclude, however, that algorithmic models of the localist connectionist variety may offer the best trade-off between clarity/transparency and formality/precision. It is the increased level of precision that allows localist connectionist models to achieve greater descriptive adequacy (Jacobs & Grainger, 1994) without sacrificing explanatory adequacy.

5. Apart from greater explanatory and descriptive adequacy, localist connectionist models offer a simple means of quantifying pre-existing verbal-boxological models that have already stood the test of extensive empirical research. Referring to this point, Page and Norris (1998) speak of a symbiosis between verbal theorizing and quantitative modeling. Furthermore, the principle of nested modeling has been readily applied with localist connectionist models. Adopting this approach facilitates the process of model-to-model comparison. Models differing by a single feature (e.g., interactivity, Jacobs & Grainger, 1992), can be compared, and different variants of the model can compete in strong inference studies (e.g., Dijkstra & van Heuven, 1998).

6. Finally, localist connectionist models, using the same simple processing units and activation functions, provide a unified explanation for phenomena observed in the different subdomains of human cognition. The general principles that govern processing in all localist models (e.g., similarity based parallel activation, lateral inhibition) can also be isolated and analyzed in an easily interpretable manner (see e.g., Grainger & Jacobs, in press). We therefore conclude that localist connectionism provides an excellent starting point for the development of a unified theory of human cognition.


Dijkstra, T. & van Heuven, W.J.B. (1998). The BIA model and bilingual word recognition. In J. Grainger & A.M. Jacobs (Eds.), Localist connectionist approaches to human cognition. Mahwah, NJ.: Erlbaum.

Grainger, J. & Jacobs, A.M. (1998). On localist connectionism and psychological science. In J. Grainger & A.M. Jacobs (Eds.), Localist connectionist approaches to human cognition. Mahwah, NJ.: Erlbaum.

Grainger, J. & Jacobs, A.M. (1998). Temporal integration of information in orthographic priming. Visual Cognition, in press.

Green, CD. (1998) Are Connectionist Models Theories of Cognition? PSYCOLOQUY 9(4)

Jacobs, A.M. & Grainger, J. (1992). Testing a semistochastic variant of the interactive activation model in different word recognition experiments. Journal of Experimental Psychology: Human Perception and Performance, 18, 1174-1188.

Jacobs, A. M., & Grainger, J. (1994). Models of visual word recognition: Sampling the state of the art. Journal of Experimental Psychology: Human Perception and Performance, 20, 1311-1334.

Jacobs, A.M., Rey, A., Ziegler, J.C, & Grainger, J. (1998). MROM-P: An interactive activation, multiple read-out model of orthographic and phonological processes in visual word recognition. In J. Grainger & A.M. Jacobs (Eds.), Localist connectionist approaches to human cognition. Mahwah, NJ.: Erlbaum.

McClelland, J. L. & Rumelhart, D. E. (1981). An interactive activation model of context effects in letter perception: Part I. An account of basic findings. Psychological Review, 88, 375-407.

Page, M. & Norris, D. (1998). Modeling immediate serial recall with a localist implementation of the primacy model. In J. Grainger & A.M. Jacobs (Eds.), Localist connectionist approaches to human cognition. Mahwah, NJ.: Erlbaum.

Rumelhart, D.E., Hinton, G.E. & Williams, R.J. (1986). Learning internal represenatations by error propagation. In D.E. Rumelhart, J.L. McClelland, & the PDP research group, Parallel distributed processing: Explorations in the microstructure of cognition (Vol. 1). Cambridge, MA: Bradford Books.

Volume: 9 (next, prev) Issue: 10 (next, prev) Article: 7 (next prev first) Alternate versions: ASCII Summary