1. Last time we talked about the possibility that a connectionist architecture of the mind might offer an alternative architecture to the classical or symbolist architecture. One way to think about this is that a connectionist architecture might solve ecological level problems, but without a computational or algorithmic level description. (One could describe the working of the network that way, but the network itself would not have parts that were isomorphic to the symbols of the description.)
2. Today I want to review a few arguments that are often thought to favor connectionism.
(a) Connectionist networks look like networks of neurons, therefore connectionism is more biological plausible.
There is some controversy over the extent of any isomorphism between connectionist ‘nodes’ and brain states, but I think few wish to defend the claim that neuronal cells are very isomorphic close to connectionist nodes.
On the other hand, connectionism has succeeded in generating lots of discussion about how implementation might occur. It thus has challenged the functionalist idea that implementation doesn’t matter, and it has brought progress in understanding ways in which complex systems might interact to allow large scale computations.
(b) Connectionist networks are fast, but classical programs are slow.
This is sometimes called the 100 step argument, for the following reason. A person can make many complex judgments within a second. But a neuron can only change states, say, 100-1000 times in a second. So, whatever allows us to make judgments, it must take only 100-1000 steps. Standard classical AI programs take many more steps.
There are two problems with this argument. First, the number of steps one takes depends on the complexity of the step. For example, when you have a document on your computer, you might move the mouse to press the ‘print’ button at the top of the screen, thereby printing the document. Is that one step? Or is it many, for the computer must first print the first letter of your document, and then the second letter, and so on. The number of steps required depends on degree of complexity of the steps.
A second problem with this argument is that it seems to assume that all classical architectures must compute in a serial, or ‘one at a time’ fashion. Connectionist networks, in contrast, engage in parallel computing, performing many operations at once. But one could build a symbolist architecture that computes in a parallel fashion as well. Thus, there’s no reason to think that classical architectures are really so much slower than connectionist architectures.
On a final note: connectionist architectures, some critics point out, are not very fast learners. It takes an enormous amount of time to ‘train one up’ to perform distinguishing tasks. One might then complain that connectionist networks aren’t really so fast at all…
c. Connectionist
architectures also exhibit graceful degradatiom.
Graceful degradation of a system occurs when a system adapts to a loss of a part ‘gracefully’. That is, it keeps running, though not as well. Economies exhibit graceful degradation in response to local destruction.
The mind, it is said, also exhibits graceful degradation. One might wonder if this is the case, as there are many examples of specific deficits in the cognitive neuroscience literature. For example, prosopagnosiacs lose their ability to recognize faces. Is this graceful degradation (because other parts of their mind continue to work), or is it nongraceful, because one capacity is wiped out? Some additional clarification would be useful here.
Sterelny discusses a number of other arguments as well.
3. The real question we need to ask is whether connectionism ‘eliminates’ the need for a computational description (and thus is a rival of computational accounts), or whether it is simply a description of how symbolist architectures might be implemented in the brain.