Understanding Intelligence

Working through in psychoanalysis, adding behavioral homework to abstract cognitive schemata, enactment in a host of therapies: all require action, motor activity, and engagement with the real world. How can our psychiatric theories move from the static to the interactive? Perhaps we are daunted by what seems like a greater complexity than the universe of a disembodied neurotransmitter-stuffed skull on a shelf.

But here comes a book from artificial intelligence that says progress will be made in understanding the mind only when we add movement and engagement, which will result not in complication but in simplification and clarity of our models. Wondering about your model of a mental process? Put legs or wheels on it, give it sensors and tropisms with which to interact with the world, and learn from what it does. But how, you ask, can one put wheels on something like “selfobject” theory? Meet the “timid” Braitenberg vehicles in chapter 6. I once saw a Braitenberg vehicle described at Brown University that did a pretty good job of relating to its object in a borderline way.

Perhaps the miraculous mechanism of human vision interests you (as it does Steven Pinker [1]). Neural networks have modeled the retina, but we have no model of general vision. This book suggests that we will not have one, because we do not see with general vision with our own eyes. The puzzle of vision is simplified by considering it as the active engagement with the world that it truly is. We do not wait for the world to impinge upon us, hoping to extract forms by inner-outer matching. Instead, we actively search, target, and focus our visual sampling with a simplifying purpose.

The concept of general intelligence (Spearman’s factor g) is similarly suspect. The authors favor multiple intelligences. Although Duncan et al. (2) have located a center in the lateral frontal cortex that activates during high-g and low-g spatial, verbal, and perceptuomotor tasks, Sternberg (3) has argued that localization is like saying we understand a computer because we have located its chip. Markman (4), reviewing this book, acknowledged that the “perceptual and motor abilities of the new generation of robots are impressive” but stated that “it has not been shown that this kind of sensorimotor coordination will provide a better framework for understanding high-level intelligence than the techniques currently being pursued in cognitive science.” Perceptual and motor abilities are essential prerequisites for intelligence to develop and function, say Pfeifer and Scheier, pointing out that mental rotation of objects, for example, often seen as the epitome of a high-level cognitive task, is actually a motor act.

The chapter on memory is both interesting and a tease because it remains so tentative. It demolishes the traditional “storehouse” concept of memory as a limiting metaphor. Memory is not located in the individual but manifests itself in the individual’s interaction with the environment. There are methodological limitations to recording from multiple sites in a behaving animal, but the authors propose that interactive memory processes can be studied by means of autonomous (robotic) agents. Although much remains speculative at this stage, the authors present extensive design principles and “cases” (robotic agents) that rely on connectionist neural networks, which are explained briefly here and which I have explicated elsewhere (5) for psychiatrists. They then argue that “actually building a robot draws attention to the enormous complexity of behaving systems. It also shows us very clearly all the assumptions we are making and all the processes we are leaving out when we focus only on ‘high level’ processing, as in the cognitivistic paradigm” (p. 570). They admit here a fundamental problem, that robots have used digital microprocessors, whereas brains do not just run programs or have buses to transmit data.

The next step is neuromorphic engineering, mimicking brain structure (of which the artificial retina is the most prominent example so far). Sensory-motor chips are being designed without intervening digital processing (and used in a line-following robot in Zurich called Morpho).

The authors argue throughout the book for the usefulness of autonomous agents (robots) to understand intelligence (as embodied IQ, citing Edelman and Fuster but oddly omitting McCulloch [6]) and all other aspects of mind. At the book’s end, however, having considered communication between humans and robots (such as Koboyashi’s facial expression robots), the authors conclude that “without going into detail, what we can say at first consideration is that because of their different sensory-motor and physical setup, robots will have emotions of a completely different nature than human emotions” (p. 642) and will be evaluated by a Turing test. Although this is admirably respectful of William Cannon and the origin of human emotions in body organs, emotions are singled out unnecessarily. All models of natural sensory-motor phenomena will differ from the biological, but this is not the point. It is from the underlying logic, the dynamisms and discernible quanta, even of emotions, that we shall learn.

Other goodies the reader can learn about are the TOTE (testing-operating-testing again-exiting) architecture; Brooks’s subsumption architectures, which incrementally add task-achieving behaviors on top of each other; and action election dynamics for robust and flexible behavior in a complex dynamic world, achieved in a bottom-up way and marked by motivations, avoidances, and conflicts.

Rothchild (7) complained that the word “model” has been used as a substitute for any number of precise terms. I think he would have to agree that these are models that are both precise and true.

If we do not choose to build robots, we may still observe that the nature of our minds is active and that the study of movement can clarify many psychiatric and neurological disorders (8). Leiguarda et al. (9), employing photogrammetric motion analysis to quantify gestural slicing of bread in patients with Parkinson’s disease, progressive supranuclear palsy, or multiple system atrophy, showed that patients with high clinical atrophy scores tended to have greater spatial and temporal disruption of their pantomimed slicing movements.

This book has the clarity of exposition and logical exposition of engineering texts that our psychiatric books could well imitate.