Unfortunately, there are not many people around today who can take what Edelman has to offer in books like Topobiology (Edelman G.M., 1988. Topobiology: an Introduction to molecular embryology, Basic Books Inc., New York). As of 12/28/96 there were only 18 WWW sites that even contain the word "topobiology". As of 2/4/99 we were up to 45: here is a nice one. Why aren't there 300 artificial life researchers who pay attention to topobiology rather than just one? Here is a nice site that mentions Edelman.
Some intelligent comments on Edelman by Robert P. Pula. (a backup of this link).
The review by George Johnson of Edelman's book BRIGHT AIR, BRILLIANT FIRE: On the Matter of the Mind, gives a critque of Edelman's conflict with the currently fashionable neural network modeling approach, what Edelman calls the "instructionist paradigm" (in Neural Darwinism, page 317). The dominant approach in attempts to construct computer models of brain functions is to adopt the hypothesis that such functions can be produced through computations and algorithmic operations performed by a model neural network acting on information received from a simulated surrounding environment. Artificial Intelligence researchers and modelers of brains construct computer programs that take simulated sensory input, process it, and return an output response. The first part ofThe Engine of Reason, the Seat of the Soul: A Philosophical Journey into the Brainby Paul M. Churchland gives a nice introduction to this kind of work and why cognitive scientists are so excited by it.
For example, think of a neural network model that is learning to recognize and respond constructively and appropriately to input in the form of spoken words. The model is constructed by the programmer to accept a specific type of input (in this case, a digitally encoded representation of human speech) and return a specific type of output (say, printed text, in this case). During a training period, an initially randomly connected neural network adjusts connection strengths to allow for association of specific inputs with the correct, desired corresponding outputs. Edelman calls this an "instructionist paradigm" in the sense that the environment (in this case, the training set) determines the functional connectivity of the network. A totally plastic, changeable, responsive, and trainable network model can be said to be instructed how to form the proper connections that are best suited for the particular task. The only other source of control over the trained state of the network model is from the original network design developed by the modeler and put into the computer program that embodies the model. However, it is often the case that random connection strengths are used at the start of a training session. Edelman's conflict with this standard approach is based on the fact that such a way of constructing a functional neural network is totally non-biological. Edelman is attempting to construct models of the brain. Edelman is not satisfied with models that simply perform a certain task by what ever mechanism is convenient for the programmer. So, I am not interested in splitting hairs over the issue of what is and is not computation. However, I am interested in the issue of how much any particular neural network model uses biologically motivated assumptions.
Edelman is firmly in the camp of brain modelers who are interested in constructing models that are biologically plausible. Edelman has identified and explored the two key levels at which biological brains establish functionally correct or appropriate network connections: 1) the developmental processed during embryogenesis that establish general neuronal connectivity patterns (neural histogenesis, primary repertoire) and 2) what he calls somatic selection on neuronal variance or neuronal group selection for the secondary repertoire. Edelman rightly points out that biological neural networks do not confront the environment (or a training set) from the perspective of a simple, initially randomly connected network that is perfectly malleable and able to adapt to whatever connectivity is required to solve the task at hand. Edelman stresses the biological reality that sensory input is processed in parallel by many neural network components and that there is selection (in parallel) for the neural network components (sub-networks) that are most closely matched in their (untrained) responses to the input. These selected responding elements are then adapted by modification of the functional conectivity of neurons involved in those selected sub-systems. This selection step during formation of the "secondary repertoire" and the antecedent developmental establishment of the "primary repertoire" are what Edelman stresses as important for biological brains and as what is missing from the "instructionist paradigm". It is also important that Edelman is interested in using biologically plausible algorithms for modification of connection strenths in his neural network models.
Many folks, such as George Johnson, seem to feel that Edelman is making silly and unimportant distinctions between the nature of actual brain mechanisms and the nature of today's popular computational algorithms. Johnson is wrong. The main question is this: What is the best scientific plan to follow in our attempts to make functional models of minds? The most common answer at first was:
go back to the first Edelman page
Go to John's Book Page.
Go to John's
Home Page.