Claus Lamm (1998) Does Brain Activity-oriented Modelling Solve the Problem?. Psycoloquy: 9(19) Connectionist Explanation (16)

Volume: 9 (next, prev) Issue: 19 (next, prev) Article: 16 (next prev first) Alternate versions: ASCII Summary
PSYCOLOQUY (ISSN 1055-0143) is sponsored by the American Psychological Association (APA).
Psycoloquy 9(19): Does Brain Activity-oriented Modelling Solve the Problem?

Commentary on Green on Connectionist-Explanation

Claus Lamm
Brain Research Laboratory
Department of Psychology
University of Vienna
A-1010 Vienna


Claiming that it is not clear how many theoretical terms a connectionist model has to be built of is one of Green's (1998a) main arguments for referring to (a) a lack of correspondence of the theoretical entities of connectionist models to any type of empirical entity and (b) the resulting abundance of degrees of freedoms in the connectionist modelling of cognition. A more brain-oriented modelling approach might yield the desired theoretico-empirical mapping but it does not reduce a model's degrees of freedom.


cognition, connectionism, methodology, theory, computer modelling, epistemology
1. Green (1998a) refers to a lack of correspondence of a connectionist model's theoretical terms or entities to "anything known to exist in the cognitive domain" (para. 9). This statement is based mainly on the claim that the removal of single units of a network has few specific effects on a model's performance and that there are no specific rules regarding how many theoretical terms a model must be built of ("we might well have gotten just as good results with 325, or 327 units, or indeed with 300 or 350 units"; para. 11). This premise is also reflected in the conclusion (para. 23) where Green holds that part of their success "may be due to the huge number of additional degrees of freedom that connectionist models are afforded" by the use of "large numbers of uninterpreted units."

2. Green accordingly argues that connectionist models should only be considered as theories of cognition when their "units DO correspond to something closely related to the cognitive domain; viz., the neurons of the brain" (para. 18) or, more generally, when they are interpreted "as literal models of the brain activity that underpins cognition" (para. 20). This interposes the human brain as a hidden layer between connectionism and cognition. While one might agree that this layer makes a model's terms interpretable, it does not solve the "degrees of freedom" problem.

3. In para. 11, Green says that it would be ridiculous or almost meaningless that 326 units govern a specific activity in the cognitive domain. However, assigning these (or even more) units to neurons or "brain units" of whatever type makes this statement even more questionable. If connectionism is to model how the human brain processes information, there cannot be much doubt that many more units must be involved. Thus, modelling the brain's activity does not reduce a model's number of terms, it increases them.

4. This particularly applies when it is intended to model the activity of a "true brain" whose behaviour is not only determined by simple and standardized neurons, but also by, for example, glial cells, different types of synapses, area-specific variations in neuron-density and types of neurons, etc. Taking all these different variables into account would be a prerequisite (Crick, 1989; Orbach, 1998) for a connectionist model to be a LITERAL brain activity model -- especially when one wants to deal with the brain's implementation of cognition. Although this certainly accords with Green's intention (as also becomes evident in one of his replies, Green, 1998b), the immediate result will be an exponential explosion of the number of units/terms of a model.

5. One might argue that there is no such increase of degrees of freedom because now it is the brain that is modelled and not cognition itself. But in most cases it will still be the behaviourally observable processes during a specific cognitive task (e.g., reading, speaking) on which a model will be tested, whereas a true modelling of brain-implemented cognition could only be tested against the recorded activity of the brain doing this specific task. However, the language of brain processes is not the language of cognition (Gevins & Cutillo, 1995). Brain-activity oriented models of cognition accordingly need to "speak" both of these rather complicated languages. This could result in a "falling between the stools," one chair being the "degrees of freedom" and the other the testing against the brain's activity proper, perhaps unintentionally moving the models farther away from cognition itself.


Crick, F. (1989) The Recent Excitement About Neural Networks. Nature, 337, 129-132.

Gevins, A.S., & Cutillo, B.A. (1995) Neuroelectric Measures of Mind. In: Neocortical Dynamics And Human EEG Rhythms, ed. Nunez, P. L., Oxford University Press.

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

Green, C.D.(1998b) Lahley's Lesson Is Not Germane. Reply to Orbach on Connectionist-Explanation. PSYCOLOQUY 9 (7)

Orbach, J. (1998). Do wires model neurons? PSYCOLOQUY 9 (5) psyc.98.9.05.connectionist-explanation.2.orbach

Volume: 9 (next, prev) Issue: 19 (next, prev) Article: 16 (next prev first) Alternate versions: ASCII Summary