Claudio Gutiérrez
In an open market, competition can and often does take the form of knowledge assimilation or emulation. As far as recent economic history goes, it is clear to me that, due to entrenched historical reasons, US Inc. could not imitate Japan Inc. in its industrial "ways and means" without a major cultural revamping that most Americans will not welcome. On the other hand, it is commonplace to say (at least in the US) that Japan Inc. has developed its economy through concerted efforts to imitate US Inc. In any event, the result has not been, again for entrenched historical reasons, a simple reproduction of the US industrial complex but, on the contrary, an industrial style quite idiosyncratic and different from original US Inc. Something similar to that we could well envision, in spite of the profound differences between artificial and natural intelligence, for the process of imitation of AI Inc. relative to Brain Inc.
These reflections result from recent bibliographic research I have made for my Computers and the Mind course, created in 1981 –just one year after Simon and Newell's memorable Turing Lecture on the 30th anniversary of Artificial Intelligence– and which I have taught with great intellectual pleasure ever since.
As it is generally accepted, AI was born in 1950. That year was also the birthday of the computational paradigm for psychology. In fact, both AI and functionalistic psychology were born together, as Turing-Machine outgrowths. Nevertheless, functionalism (the idea that intelligence can be explained solely on the basis of function) and brain science (which works rather from structure to function) have traveled mostly independent roads right from the start and to this day. During the last eight years or so, however, there has been considerable interest in the topic of consciousness by brain scientists; at the same time, functionalist psychologists have begun to increase their interest in questions of "architecture." These two convergent motions have approached AI and brain science somewhat together.
If one dwells on peripheral considerations about the neural system, there is much evidence in favor of functionalism. For instance, sense organs and muscles are transducers, and internal transmission from and to them is totally uniform. On the other hand, it is clear that, in the case of the evolution of the human cortex, structure has preceded function, and not the other way around, as we will see in a moment.
First and foremost, there is an extraordinary neurological continuity in the mammal order: no especially human neuron or neurotransmitter, not even especially human architectural neural block (cortical column)! Incrementality, scalability seems to have been the norm (contrary to the AI experience). Quantity, not quality, seems to have made all the difference. All localized functions (modules), with the (partial) exception of language, were already in place in the high primates. What seem all-important are critical mass (four times more cortex from chimp to human, for example) and the resulting connectivity. There are approximately 30 billion neurons in the human cortex alone, 200 billion for the whole brain. Connectivity is incredible: each neuron connects, on the average, to 10,000 other neurons. Something comparable by far has still to occur in AI!
There are no central headquarters in the brain. Total lack of homunculus or "oval office" in the brain. The highest level of the brain's global "neuronet" turned out to be the hippocampus (a rather primitive structure, part of the temporal lobe and the limbic system). As to the "inner room" (thalamus), it turned out to be just some kind of relay point. No other interesting candidate for such post of honor. And of course, there is a high degree of parallelism in all these connections, what makes even less plausible the existence of central headquarters.
How did all come about? Complexity of the brain cannot be explained just by genetics: difference between the genomes of chimp and human is only 1% (in the DNA sequence of our about 100.000 genes). Great importance of development (a third term between nature and nurture): at the beginning (in infancy), everything got connected to everything else, so it seems. Great importance of selection: dying-off of neurons and synapses, until adolescence. Other than that, the "design" is the work of a loooonnnggg evolution, most of the time clearly suboptimal. Characteristically, multipurpose abound (serendipity of computationally-dreaded side effects). And a lot of opportunism: higher functions recruit preexisting structures for new purposes (ability to read, or to moving fast in the highway, seems to be related to the optical effect of... jumping from one tree branch to another!). And, of course, structure preceded function: random mutations multiplied by 4 the size of our cortex, giving origin to our "vacant" (associative) areas (that is to say, cortex not committed to perceptual or motor functions).
Parallelism does not necessarily mean redundancy. It is clearly the other way around: multipurpose of all circuits seems to be the norm. In fact, there is no graceful degrading in neural networks, only the appearance thereof. There is contingency support among (mainly) contiguous areas. But everything seems to be always actively involved in some distinct function or other. Famous cases: President Wilson's right-hemisphere stroke, Phinias P. Gage immense frontal lobe damage. In both these cases, the brain seemed to continue working normally, but subtle miscarries of function produced dreadful consequences for the person's (or the world's, in the case of the President) future. In fact, prefrontal lesions seem to have a tremendous relevance for the infamous "frame problem": remember McDermott's robot meditating interminably on the consequences of each of its actions!
There are many important economies in place: contrary to common myth, visual imaging is quite poor, since Nature itself is used as our preferred "hard disk" (we can always see again in case we have missed something). Visual awareness relies in eye saccades and differential acuity of retinal areas, rather than in complete appropriation of visual material. Richness of imagination (a graphics mental faculty that would permit us to project in front of us our –however poor– "natural slides" for our friends to rejoice with the images of our last vacation) is completely avoided, because of its characteristic enormous consumption of computational resources (contrast with the case of language, in which we do have both interpretative and generative functional modules: Wernike's and Broca's areas). There are no redundant discriminations either: once a feature is rendered, the consequences for behavior are drawn (no "re-presentation" for the benefit of an internal –non existent– audience!).
And finally, knowledge representation is very, very economical: in the case of colors (trichromaticity); and of tastes (four types); something similar goes for the sense of touch. There is hierarchical feature-extraction for vision (although "reentrant" or "interactive"). But "the grandmother neuron" is a myth: it has been shown that common names involve less neurons than proper names, each of which needs a loootttt of neurons to be represented. This seems to be a wonderful confirmation of Quine's claim that proper names can be replaced by variables on the basis of Russell's rendering of definite descriptions. So, we may not need inductive logic after all: general nouns seem to be primordial in our mind, not derivative and the product of tortuous construction, as Carnap assumed!