Orbach (1998) incorrectly interprets my target article (Green 1998) as claiming that connectionist networks actually model neural activity, whereas in reality I argue that nets will NEED to model neural activity if they are to model anything at all.
2. Orbach's reaction seems to be based on a misreading of the passage in question. My intent was not to press connectionists to make such a bold claim. On the contrary, in the very next paragraph (para. 19) I wrote that "the general aversion to making very strong claims about the relation between connectionist models and brain is not without good reason." The point of my target article was not that connectionists would be JUSTIFIED in claiming that connectionist networks model brain structure directly. Rather, it was suggested that there may be no clear theoretical content to their models unless they can make such a claim stick. Whether or not they can achieve this is an issue on which I am relatively agnostic.
3. Having said that, I don't know that "Lashley's Lesson" constitutes a major obstacle to the goal. Orbach tells us that "the junction at the synapse is anything but a solder-joint. It too varies from moment to moment for a number of reasons, one being the micromovements of end feet on the post-synaptic cell. The all-or-none law on which connectionist models are based has been interpreted incorrectly to mean that the neuron, regardless of its condition, either fires or does not fire to a particular excitatory stimulus" (para. 2). He is undoubtedly right in the first of these claims, but I see no reason to believe that these aspects of neural activity need elude the connectionist's modeling ability IN PRINCIPLE, even if no connectionists have thus far cared to include such details in their nets. As for the "all-or-none" law, I believe that Orbach is mistaken here. I would venture to guess that the majority of published connectionist models of cognition do not use the kind of "threshold units" that form a natural analogy to all-or-none functioning. On the contrary, they mostly use continuous activation rules.
4. In short, I am afraid Orbach missed the point of my target article, which was not that connectionist networks, as a matter of fact, directly model neural activity and that people are somehow afraid to come out and say so. It was, rather, that unless such networks can be shown to model neural activity it is not at all clear what, precisely, it is they model. I wish Orbach luck with his new book on Lashley and Hebb (Orbach, in press).
Green, CD. (1998) Are Connectionist Models Theories of Cognition? PSYCOLOQUY 9(4). ftp://ftp.princeton.edu/pub/harnad/Psycoloquy/1998.volume.9/ psyc.98.9.04.connectionist-explanation.1.green.
McClelland, J. L., Rumelhart, D. E., & Hinton, G. E. (1986) The appeal of parallel distributed processing. In: Parallel distributed processing: Explorations in the microstructure of cognition (vol. 1), ed. Rumelhart, D. E. & McClelland, J. L., MIT Press.
Orbach, J. (1998). Do wires model neurons? Commentary on Green on connectionist-explanation. PSYCOLOQUY 9 (5). ftp://ftp.princeton.edu/pub/harnad/Psycoloquy/1998.volume.9/ psyc.98.9.05.connectionist-explanation.2.orbach.
Orbach, J. (in press). The Neuropsychological Theories of Lashley and Hebb. Lanham MD: University Press of America.