Asim Roy insists (Connectionism, controllers and a brain theory) that there are controllers in the brain. This is not as sinister as it might sound.
Roy presents his views as an attack on connectionist orthodoxy. Connectionists, he says, believe that the brain does not have in it groups of cells that control other parts of the brain. He cites many sources, and quotes explicitly from Rumelhart, Hinton, and McClelland, who say
“There is one final aspect of our models which is vaguely derived from our understanding of brain functioning. This is the notion that there is no central executive overseeing the general flow of processing…”
Roy begins by addressing a startlingly radical argument against the notion of controllers in any system. He takes up the analogy of a human being driving a car. The steering and other systems, according to a normal understanding, are controlled by the driver; but, he says, there is an argument that this is the wrong way of looking at it. It’s true that the steering is guided by input from the driver, but there is also feedback to the driver from the various systems of the car, which also determine what the driver does. The current position of the wheel and the response of the car determine how much force the driver applies to the wheel. So really the car is controlling the driver as well as vice versa, and the very idea of a controller within a system is a misapprehension.
This is obviously a slightly silly argument, but it serves to show that we need to define the notion of a controller properly. Roy suggests that the essence of a controller is that it is not dependent on inputs. The steering moves only in response to my inputs, whereas if I choose (and am tired of life, presumably) I can ignore any feedback from the car and turn the wheel any arbitrary number of degrees I settle on. Similarly, although my channel-hopping is normally guided by what appears on the TV screen, if I wish I can choose to change channels arbitrarily, or tap out a rhythm on the remote control which has nothing to do with the TV. A controller, in short can operate in different modes while a subservient system cannot.Armed with this definition, Roy argues that a connectionist network which learns by back-propagation requires an external agent to set the parameters, whether it be a human operator or another module within the overall system. I suppose it could be retorted that this controller is indeed external, and exercises its influence only during learning, but Roy would probably say that if we’re modelling the human brain, these controllers would have to be taken to be part of the neural set-up. In any case, he goes on to give examples from neuroscience to show that some parts of the brain do indeed seem to operate as controllers of some other parts. I must say that this claim seems so evidently true to me that argument in its favour seems almost redundant.
Roy’s conclusion is that his reasoning opens the way to a better approach, where instead of being left to local learning, the methods and weightings to be used can be dictated by a central system, at least on some occasions.
I think Roy’s conclusion that there must be controllers within the brain is hard to disagree with: the question is more whether he’s demolishing a straw man. What did Rumelhart, Hinton, and McClelland actually mean? I suspect they meant to deny the existence of a central ‘homunculus’. a little man in the brain who does all the real work; and also to deny that the brain has a CPU, a place where all the data and instructions get matched together and processed. I don’t really think they meant to deny that any part of the brain ever controls any other part. I’m not sure that connectionists have ever reached the point of proposing an overall architecture for the brain, or even that that would be within the scope of the theory; rather, they just want to investigate a way of working which may be characteristic of parts of the brain.
I can imagine two ways connectionists might respond to Roy’s claims without directly contesting them. One would be to accept that the control function he describes exists, but claim that it doesn’t reside in one fixed place. Different parts of the overall network might be in control at different times; it’s not that bits of the brain don’t control other parts, merely that control is sort of smeared around. But equally I can imagine a connectionist simply saying; of course we never meant to deny that that sort of control relation exists, so thanks for the clarification…
I think Roy’s attempt to define what a controller does is interesting, however. A controller, on his view, can follow the inputs, or operate without them. But surely other systems can operate without inputs, too? If I black out and cease to provide the car with control inputs, it doesn’t cease to function. It may function disastrously, but that’s also true if I exercise my controller’s right to ignore inputs and take a kind of existentialist approach to steering. You could say that in cases like the black-out one the controlled system is continuing to receive inputs – they’re just consistently zero. The point really is not operating without inputs, but being able to ignore the ones you’re getting. I think what we’re grasping for here is that the inputs to a controller don’t determine the outputs. That’s interesting because it sounds like one version of free will. When I act freely, my actions were not determined by the environmental inputs. But it’s notoriously hard to explain how that could be so in a deterministic world.
If you’re a softy compatibilist about free will, you may be inclined to argue that it is to some extent a matter of degree. Where input A always gets output B, no question of freedom or controllerhood arises. But if there is a complex internal algorithm working away, such that input A may get any letter of the alphabet on different occasions, things start to look different. And if the outputs begin to show a certain kind of coherence or meaningful salience – if they begin to spell intelligible words – we might be inclined to say that the system is in some sense in control. If when we turn the wheel, the car does not respond, we might gasp metaphorically “The damn thing’s got a will of it’s own!”; further, if it actually directs itself down a side road and into the filling station we might seriously and literally begin to credit the car with intelligence. So far so Dennettian.
If that line of reasoning is right, controllerhood is indeed a tricky business: it might be that the only way to know whether some group of cells is a controller would be to watch it and see whether it did controllery things. And if that’s the case, maybe smeary connectionism has something in it after all…