Embodied Cognition

Until recently, Artificial Intelligence research has been grounded on a theory of cognition that is based on symbolic reasoning.  That is, somewhere in our heads the concepts are represented symbolically and reasoned about via deduction and induction.

At long last, AI researchers are truly learning from human cognition (oh the irony!)  Introducing Leo, the robot that learns to model and reason about the world like human babies do, via embodied experience and social interactions:

“It’s really through the body, and the dynamic coupling of neural systems for perception, action and introspection, that cognition emerges,” says developmental psychologist Linda Smith of Indiana University in Bloomington.

Smith goes even further in challenging the conventional wisdom on human intelligence:

“That’s all there is to cognition,” Smith somewhat defiantly told an audience at the cognitive science meeting. Symbolic representations of knowledge in the brain, cherished by many cognitive scientists, simply don’t exist, in her view.

Her view is supported not only by experimental results in infants, but also by vast amounts of cognitive science literature on the embodied nature of cognition.  Lakoff and Johnson’s tour de force, Philosophy in the Flesh, summarizes these results and presents a theory of cognition based on the embodied mind.  They contend that the primary mechanism by which conscious reasoning is done is via metaphor, primarily metaphor that is based on the five senses (e.g. “I see what you mean”, or “That was a bittersweet experience”).

What’s novel about Smith et al’s work is that they bring in the social dimension to cognition and how powerful the results appear to be.  One way to extend their approach would be to endow Leo with the six social influence primatives catalogued by Cialdini: Reciprocation, Commitment & Consistency, Social Proof, Authority, Liking, and Scarcity.  Currently Leo’s main cognitive trick seems to revolve around Social Proof.

  • Pingback: A Digital Outrigger / Library Usability Links 12/20/08()

  • valter

    I read “Philosophy in the flesh” when it came out and I had the feeling the authors were: 1) probably right; 2) overstating the evidence in their favour.

    How has the debate moved on since then?

  • @ valter

    Interesting statement/question. I’m not sure if you really are saying that they are overstating evidence or simply “methinks thou dost protest too much”. Assuming the latter I would say that I disagree with your point (2) in the sense that I don’t think that most people agree with — or at least fully understand the implications of — (1). And since they don’t, the case needed (still needs) to be made ad nauseum.

    If you feel they are truly overstating evidence, I guess it would be good if you can characterize that a bit further. I was pretty well versed in the primary cognitive science research of the day and I didn’t detect any exaggeration or falsehood. On the contrary, I found myself nodding in appreciation and agreement as they synthesized the data into a coherent and self-evidently correct model of cognition that answered questions that competing models of the day did not. Which is not to say that Lakoff and Johnson’s theories are complete.

    So in terms of how the debate has moved on, I would say not very much, unfortunately. But the reason for this post was that as of late, AI researchers seem to be catching on and I highlight one of the few promising examples.

  • Valter


    Thanks for the reply.

    I meant it as “truly overstating the evidence” (e.g., using experiments in favor of metaphor-based embodied cognition for some relatively simple linguistic processes and then claiming that practically all cognition works in the same way).

    But I am only an amateur with no direct knowledge of the primary literature.

    And I have not yet read Lakoff’s more recent book (with Nunez) on the cognitive basis for mathematics. Btw, any thoughts about that?

  • Haven’t read it yet, but I just put it on my list, thanks :-)

  • Valter

    btw, you may also be interested in Keith Devlin’s recent columns on the MAA website (http://www.maa.org/devlin/devlin_12_08.html and http://www.maa.org/devlin/devlin_01_09.html).

    For example, he writes:

    “In particular, I provided evidence in support of my thesis (advanced by others in addition to myself) that, whereas numbers and perhaps other elements of basic, K-8 mathematics are abstracted from everyday experience, more advanced parts of the subject are created and learned as rule-specified, and often initially meaningless, “symbol games.” The former can be learned by the formation of a real-world-grounded chain of cognitive metaphors that at each stage provide an understanding of the new in terms of what is already familiar. The latter must be learned in much the same way we learn to play chess: first merely following the rules, with little comprehension, then, with practice, reaching a level of play where meaning and understanding emerge. (Lakoff and Nunez describe the former process in their book Where Mathematics Comes From. Most of us can recall that the latter was the way that we learned calculus - an observation that appears to run counter to - and which I think actually does refute - Lakoff and Nunez’s claim that the metaphor-construction process they describe yields all of pure mathematics.)”

  • While I hesitate to comment on something I haven’t read yet, I’ll suggest that Devlin’s conclusion may be a bit off the mark. I would guess that Lakoff and Nunez would say that the primitive cognitive operator that underlies all symbolic reasoning (which includes all rule-following) is in fact the mechanism of metaphor-construction. Which is not much more complicated than spreading activation coupled with the canalization of co-variant neural firing patterns. Of course there are layers of executive function and heuristics on top of this primitive operator and it’s definitely worthwhile examining these. But it’s not hard to see how metaphor-construction can be used to bootstrap modus ponens (i.e. deductive logic) and also inductive reasoning. Thus when Devlin claims that metaphor-construction doesn’t/can’t yield all of mathematics, I suspect in his mind he’s not thinking about this but rather just the grounding of the basic metaphors into embodied constructs. If you use the more general definition of metaphor-construction, then there is no conflict between the two viewpoints.

    As a somewhat relevant aside, I actually took predicate calculus from Devlin when I was in college, and though I eventually “got it”, had a hard time learning the way it was being taught. The mistake that both camps tend to make (and I would throw Chomsky in with Devlin’s camp) is that they over-generalize and insist that we are all fundamentally learning the same material via the same cognitive mechanisms. Well, maybe that’s a little unfair. But I do think we need to look into the various “styles” of learning that educators talk about and see whether they map consistently to a set of (possibly composite) cognitive operators that each of use relies on in different mixture and degree. I suspect this to be the case.

  • Might I suggest, my book 2001 book, The Raft, which seems to back up Rafe Furt’s perspective both in the original post and the follow-ups?


    I’ve begun publishing an “open text” version to bring more people into the discussion, like this.

    John Micahel Vore