Andy Clark on extended mind, A.I., and predictive processing

I want to point to a New Yorker article by Larissa MacFarquhar describing the evolution of the ideas of philosopher of mind Andy Clark.

The first section of the article follows Clark's development of the idea that our minds must be defined as extended beyond our bodies to include the tools in our environment without which they cannot function:

Clark started musing about the ways in which even adult thought was often scaffolded by things outside the head. There were many kinds of thinking that weren’t possible without a pen and paper, or the digital equivalent—complex mathematical calculations, for instance. Writing prose was usually a matter of looping back and forth between screen or paper and mind: writing something down, reading it over, thinking again, writing again. The process of drawing a picture was similar. The more he thought about these examples, the more it seemed to him that to call such external devices “scaffolding” was to underestimate their importance. They were, in fact, integral components of certain kinds of thought. And so, if thinking extended outside the brain, then the mind did, too.

It then describes his moving into artificial intelligence and robotics, encountering the work of Rodney Brooks at M.I.T:

Maybe the way to go was building an intelligence that developed gradually, as in children—seeing and walking first. Perhaps intelligence of many kinds, even the sort that solved theorems and played chess, emerged from the most basic skills—perception, motor control...While constructing a robot that he called Allen, Brooks decided that the best way to build its cognition box was to scrap it altogether. ...It was controlled by three objectives—avoid obstacles, wander randomly, seek distance—layered in a hierarchy, such that the higher could override the lower...It would make no plans. It would simply encounter the world and react.

Robots like Allen... seemed to Clark to represent a fundamentally different idea of the mind. Watching them fumble about, pursuing their simple missions, he recognized that cognition was not the dictates of a high-level central planner perched in a skull cockpit, directing the activities of the body below. Central planning was too cumbersome, too slow to respond to the body’s emergencies. Cognition was a network of partly independent tricks and strategies that had evolved one by one to address various bodily needs. Movement, even in A.I., was not just a lower, practical function that could be grafted, at a later stage, onto abstract reason. The line between action and thought was more blurry than it seemed. A creature didn’t think in order to move: it just moved, and by moving it discovered the world that then formed the content of its thoughts.

Then, how does the brain make sense of the world?

To some people, perception—the transmitting of all the sensory noise from the world—seemed the natural boundary between world and mind. Clark had already questioned this boundary with his theory of the extended mind. Then, in the early aughts, he heard about a theory of perception that seemed to him to describe how the mind, even as conventionally understood, did not stay passively distant from the world but reached out into it. It was called predictive processing.

It appeared that the brain had ideas of its own about what the world was like, and what made sense and what didn’t, and those ideas could override what the eyes (and other sensory organs) were telling it. Perception did not, then, simply work from the bottom up; it worked first from the top down. What you saw was not just a signal from the eye, say, but a combination of that signal and the brain’s own ideas about what it expected to see, and sometimes the brain’s expectations took over altogether.

One major difficulty with perception, Clark realized, was that there was far too much sensory signal continuously coming in to assimilate it all. The mind had to choose. And it was not in the business of gathering data for its own sake: the original point of perceiving the world was to help a creature survive in it. For the purpose of survival, what was needed was not a complete picture of the world but a useful one—one that guided action. A brain needed to know whether something was normal or strange, helpful or dangerous. The brain had to infer all that, and it had to do it very quickly, or its body would die—fall into a hole, walk into a fire, be eaten.

So what did the brain do? It focussed on the most urgent or worrying or puzzling facts: those which indicated something unexpected. Instead of taking in a whole scene afresh each moment, as if it had never encountered anything like it before, the brain focussed on the news: what was different, what had changed, what it didn’t expect...This process was not only fast but also cheap—it saved on neural bandwidth, because it took on only the information it needed—which made sense from the point of view of a creature trying to survive...To Clark, predictive processing described how mind, body, and world were continuously interacting, in a way that was mostly so fluid and smoothly synchronized as to remain unconscious.

And, summarizing paragraphs,

He knew that the roboticist Rodney Brooks had recently begun to question a core assumption of the whole A.I. project: that minds could be built of machines. Brooks speculated that one of the reasons A.I. systems and robots appeared to hit a ceiling at a certain level of complexity was that they were built of the wrong stuff—that maybe the fact that robots were not flesh made more of a difference than he’d realized. Clark couldn’t decide what he thought about this. On the one hand, he was no longer a machine functionalist, exactly: he no longer believed that the mind was just a kind of software that could run on hardware of various sorts. On the other hand, he didn’t believe, and didn’t want to believe, that a mind could be constructed only out of soft biological tissue. He was too committed to the idea of the extended mind—to the prospect of brain-machine combinations, to the glorious cyborg future—to give it up.

In a way, though, the structure of the brain itself had some of the qualities that attracted him to the extended-mind view in the first place: it was not one indivisible thing but millions of quasi-independent things, which worked seamlessly together while each had a kind of existence of its own. “There’s something very interesting about life,” Clark says, “which is that we do seem to be built of system upon system upon system. The smallest systems are the individual cells, which have an awful lot of their own little intelligence, if you like—they take care of themselves, they have their own things to do. Maybe there’s a great flexibility in being built out of all these little bits of stuff that have their own capacities to protect and organize themselves. I’ve become more and more open to the idea that some of the fundamental features of life really are important to understanding how our mind is possible. I didn’t use to think that. I used to think that you could start about halfway up and get everything you needed.”