I have been hearing a lot of buzz lately about Asim Roy’s newly published paper Connectionism, controllers, and a brain theory. What is all the hype about? Well, Roy claims to be offering a “new theory for the internal mechanisms of the brain.” Sounds exciting doesn’t it? What could Roy be proposing that is so revolutionary? Roy proposes that…wait for it… some parts of the brain control other parts! I suppose most of you aren’t exactly blown away, and I certainly wasn’t either.
Roy’s rhetoric is obviously pretty overblown, but let us give him the benefit of the doubt and move on to his actual arguments. He starts off by claiming that connectionist theory “postulates that the brain does not have controllers in it.” He quotes Rumelhart, Hinton, and McClelland as saying “there is no central executive overseeing the general flow of processing.” Seems pretty non-controversial to me though. They seem to be merely saying that there isn’t a homunculus in the system, controlling everything with a immutable Will. Philosophy 101. So what is Roy actually arguing against? A straw man? Sort of, but not quite. Roy argues that in connectionist models, there is a “controller” in the system that controls the learning algorithms and thus connectionist theories are essentially rooted in control-theoretic modeling.
But, as peter over at conscious entities mentioned, connectionist theorists haven’t exactly gotten to the point where they are proposing a general architectural model of how the brain works. It seems entirely plausible that when connectionist models get to that point of complexity, they wouldn’t hesitate to propose that some modules control other modules. Otherwise, I don’t see how one could get a theory that modeled high-level cognition. The way Roy structured his arguments, I don’t think anyone would argue against the idea that “there are parts of the brain that control other parts.” Furthermore, Roy himself undermines his claim for proposing a “new paradigm” when he says things as trivially obvious as:
It should be pointed out that this theory does not posit that there is a single executive controller in the brain. [b]Instead it envisions “multiple distributed controllers” controlling various subsystems or modules of the brain[/b]. The main argument of the paper that connectionists use “executive controllers” is only pointing out that their algorithms use a “central controller.” But different modules in the brain using connectionist-type learning can have their separate controllers.
I’d also like to point out that Roy was beaten by at least ten years on his emphasis of controllers. In his 1997 book Being There, Andy Clark says:
The idea here is that the brain should not be seen as primarily a locus of inner descriptions of external states of affairs; rather, it should be seen as a locus of inner structures that act as operators upon the world via their role in determining actions…This perspective leads to a rather profound shift in how we think about mind and cognition-a shift I characterize as the transition from models of representation as mirroring or encoding to models of representation as control
So contrary to Roy’s strong rhetoric, people sympathetic to connectionist theory such as Clark have been thinking about the mind and the brain in terms of action-oriented controllers for many years. In conclusion, I agree with Roy’s essential argument that there are parts of the brain that control other parts of the brain, but I don’t think this is a revolutionary of a paradigm as he thinks it is. Roy himself quotes from all over the neuroscience literature showing that it is riddled with control-theoretic terms, and by his own argument, he shows that connectionist theory is also already steeped in control theory. Surely, the connectionists themselves understand this. So who is Roy arguing against here?