Report on research of karl macdorman

Candidate for Tenure, IU

I was one of the external examiners for Karl MacDorman’s Cambridge University PhD in 1997 because his thesis was on the topic of symbol grounding, cognition and robotics. I will now try to put my understanding of his subsequent work into the framework of their significance for research progress on this problem.

The research program itself has two aspects, the more general and ambitious one is (1) to design robots that can do what people can do, partly in order to provide useful resources for people, and partly to explain how people themselves function. The latter is the gist of the Turing Test (TT), which is that the way to explain how people function is to design robotic systems that can do what people can do. Once the performance capacities of those systems approach human scale – so much so that people can no longer tell apart the robots from the people (for a lifetime, in principle) – then chances are that the functional mechanism underlying their performance is the same.

But we are still very far from designing robots that can pass the TT. All they can do is tiny fragments of what people can do. Most of these arbitrary fragments of our performance capacity can be generated in many different ways, hence it is unclear which if any of those ways would scale up to full TT scale. This is where the other aspect of macDorman’s research program comes in (2) to design robots that can use and understand natural language: Natural language is more than an arbitrary fragment of human performance capacity. It is a lifesize chunk that encompasses many of our capacities. To be able to speak and understand language, you need to be able to identify and name and manipulate all the things that the words in the language (symbols) stand for. The natural candidate mechanism for this capacity is the power of computation (also demonstrated by Turing), but computation has a problem: Computation is the manipulation of symbols using rules (algorithms) that are based on the arbitrary shapes of the symbols (syntax), not on their meanings (semantics). In mathematics, logic and computation, it does not matter that it is all syntactic, but in robotics and cognition it matters, because the words of natural language do have meanings, and we manipulate them on the basis of their meaning, not just their shapes.

Part (but not all) of meaning is reference: How to connect the internal symbols of people and robots with the things in the world that the symbols refer to. MacDorman’s thesis addressed this by examining the process of sensorimotor development: Symbols must get connected to their referents through sensorimotor structures and learning. One of the problems of the original Turing Test was that it was purely verbal: Human interactions with the artificial system were only via what we would call email today. This leaves out all sensorimotor interactions and restricts the test to computation, which is ungrounded. The reason Turing had excluded sensorimotor interactions is doubly relevant to MacDorman’s later work. Turing had rightly reasoned that in testing whether or not people could distinguish the artificial system’s capacities from those of real people, the appearance of the system should not be allowed to bias people’s judgment. (In those days, there were not yet robots that could look or act convincingly human.) So the original TT was restricted to verbal interactions. MacDorman’s work rightly opens it up to testing whether the system can perform distinguishably from a human also in its sensorimotor capacities: Can it recognize and manipulate objects, not just words, and can it connect the words to the right objects?

But there is also another aspect of appearance, over and above the capacity for sensorimotor interactions with real-world objects that are the referents of our words, and that is appearance itself, as a form of animate structure and motion (of the face, voice and body) that we can all perceive and produce. The goal of the TT is to design a robot that is completely indistinguishable in its performance capacity from a human being, and that includes everything it can do with its body, hence, to an extent, it also includes the “shape” of its body. It turns out that people have remarkably acute capacities for detecting other people’s internal states based on viewing their external behavioral manifestations. Developmental psychologists have come to call these or “mind-reading” capacities – not in the sense of telepathy or other “psychic” powers, but in the sense of being able to infer, from people’s appearance and behavior, what they think, feel, know and want. Part of this mind-reading capacity is our uncanny ability to detect when something is being done mechanically, by a machine, rather than by a thinking, feeling human being.

Our mind-reading capacities are not infallible: We can sometimes mistake a machine for a person, or a person for a machine, for a while. But in the long run, it will take nothing short of a TT-scale robot to “fool” us -- and at that point, as Turing points out, it is no longer clear that we would be wrong in thinking the robot really has mental states.

Karl MacDorman’s research program is bringing us closer to that day. Beyond the question of how robots could ground their symbols and talk with us intelligibly while interacting with the same external world with their senses and their movements, MacDorman is studying what aspects of their appearance underlies our perceiving them as being feeling creatures like ourselves, rather than just mechanical devices. The “uncanny valley” work is looking not only at what properties will make robots more convincingly lifelike for us, but also at what are the “edges” of our detectors, the homologues of our optical illusions, in which certain combinations of features make us perceive a robot or dynamic image as “eerie.” This work helps bring us closer not only to generating the real thing (for cognitive science), but also to generating more convincing ways to fool us (for the entertainment industry). MacDorman’s work on “android science” also shows the reverse side of Turing Testing: Not only is there interest in using humans to test whether robots can do what humans can do well enough to be indistingusihable to humans, but there is interest in using convincing androids to test human perceptual capacities and biasses, espscially in social perception.

MacDorman;s work on cultural differences in attitudes toward robots is also of interest, showing how differences in people’s experience, values and expectations affect their judgments and pereception. His work on computational prediction of human emotional response to music is a natural extension of this work, and his findings on video surveillance techniques are also of obvious relevance today in the complementary area of creating useful artificial resources.

Overall, I find that, in comparing MacDorman with other candidates for tenure, he has a productive, timely research program and that his findings to date, as well as the promise of his research program, are of sufficient practical, interdisciplinary and international scope to make him a credit to Indiana University’s many existing strengths in cognitive science.