In his recent novel Galatea 2.2, Richard Powers has his protagonist (also called Richard Powers) reminisce about the time when he used to mull over questions of cognition and the mind. "What was memory? Where, if anywhere, did it reside? How did an idea look? Why was comprehension bred, or aesthetic taste, or temperament? Predicates threaded my neural maze. After great inference, I came to the conclusion that I hadn't the foggiest idea what cognition was. Nobody did, and there seemed little prospect of that changing soon. No tougher question existed. No other, either. If we knew the world only through synapses, how could we know the synapse? A brain tangled enough to tackle itself must be too tangled to tackle." Then, on the cusp where the novelist's past yields to the present and his plot kicks into full gear, he encounters a way out of the recursive deadlock, a way that allows Richard to talk about Richard without getting lost in an infinite regress. The magic word that opens the door from the vicious circle of self-reflection into the mind's brave new world is -- connectionism. In between the dry Artificial Intelligence symbolists and the neurochemical wetware workers the connectionists propose a new picture of the brain: no longer a slogging information processor, the brain is now seen as a creative modelmaker whose most important operations occur in the connections between its components (neurons). The brain does its work through large, multilayered neural networks consisting of some hundred billion neurons that associate, learn, and judge without being cognizant themselves. Conscious intelligence is an illusion, logic nothing but massively parallel pattern matching. We humans only pretend to be syllogistic creatures. While we're thinking that we're thinking, our synapses are in fact firing to set off the associative matrix corresponding to an input (a premise, say), and while it feels as if we are deducing something, nothing but electrical impulses diffuse through webs of neurons. When two synapses fire simultaneously, Hebb's law makes their connection grow stronger as a reward. This is called learning by reinforcement. But a net can be trained also by error correction: if the output (the conclusion, say) differs from the result desired by the trainer, the error is propagated back through the net, not via the neuronal paths (which are unidirectional) but through higher-level brain processes (call it the executive suite) that have the authority to adjust the firing conditions and reduce the connections that contributed to the error in the first round. Richard learns from Richard's hindsight. With this thumbnail sketch, we've summarized the middle part of Artificial Minds. Professor Franklin is "a mathematician, turned computer scientist, turning cognitive scientist." He promises to be our tour guide to the greatest monuments cognitive scientists have erected over the past twenty years. But from the start he also sets down the assumptions we tourists have to adopt in order to enjoy the trip. There is, first, the physicalist or materialist hypothesis: thought (whatever it is) arises from neural firings; or, mind is what brain does. In "functional" terms, mental activities play on a physical substructure, similar to a compact disc making music on a CD player. Secondly, when sighting a mind, our guide will point out only the mind's "mechanisms," how mind works and how it displays its measurable behavior. As soon as the mind turns inward upon itself and reveals what poets would call the human soul, the author turns away. Souls have no mechanisms. Thirdly, before we embark, we should agree on the purpose of mind: the mind determines "how to do the next thing" and how to do it right -- which translates into maximizing pleasure and minimizing pain for the owner of said mind. But there's an even more important presupposition. It's an engineer's ver sion of a cognitive principle that is epitomized in a quote by Carver Mead, neuro-computer scientist extraordinaire, whom Franklin calls "one of my heroes": "If we really understand a system we will be able to build it. Conversely, we can be sure that we do not fully understand the system until we have synthesized and demonstrated a working model." One might argue that a good designer understands his machine without building it; that a literature professor can understand Donne's poetry without writing holy sonnets herself; that a mathematician can appreciate the proof to Gšdel's Incompleteness Theorem without having a clue how Gšdel ever got his idea. Does an aerospace engineer understand birds? But we shall not argue yet, because Mead's quote is not a proposition but an operational definition: building is understanding, and vice versa, basta cosi. Agreeing to this axiom is essential for the tour; it is the justification for cognitive researchers who strive to understand natural minds by building artificial minds. With these presumptions duly stated, the trip beings. Our guide has parsed the problem of researching "mechanisms of the mind" into "three fundamental questions" that he'll pose to every artificial mind encountered on the way: 1. What constitutes "cognition" for this mind? 2. How does cognition work? 3. How do I know when mind is functioning adequately? The first stop, at the monuments of symbolic Artificial Intelligence (AI), has attracted as many early disciples as it has provoked strong detractors in recent years. AI's contention that cognition is, or can be reduced to, symbolic computation, was attractive to electronics enthusiasts in the seventies. They were intrigued by the striking analogies between brain and hardware, mind and software. When their rule-based symbol manipulations solved a (usually algorithmic) problem, they believed some mindful work had been accomplished. The scoffers, from the Dreyfus brothers and John Searle in the halcyon days of AI to Roger Penrose (The Emperor's New Mind [1989]), have pointed out that symbolic processing can never represent our common sense, our adaptability, our native competence (Dreyfus); that symbol manipulation has nothing to do with understanding, for instance, the meaning in a foreign language (Searle). The most sophisticated, and the most speculative, attack comes from Penrose. His argument depends on Gšdel's Incompleteness Theorem (roughly: "every sufficiently powerful and consistent formal theory allows us to state propositions that are true but unprovable within that theory"). The human mind can divine (mathematical) truths that no algorithm (rules backed by the theory) can ever prove; ergo, mind is not symbolic computation. Instead, Penrose speculates that deep inside the neurons, where the microtubules reside, subtle quantum effects consort to build our consciousness. Franklin has his own ideas. Going a step further, Franklin rejects the AI symbolists' idea that rules govern brain function; he also diminishes the connectionists' image of mind as modelmaker, "representing" the outside world through patterns. The scent of a perfume is not represented in any pattern in the olfactory cortex. Instead (this is the crux for Franklin), certain "autonomous agents" create the sensations on the fly as they manufacture information out of sense stimuli. The activities of these agents, "with no human in the loop," is what mind does, to wit: "The process by which an autonomous agent selects actions, including those processes which by objectivation, categorization, and so on create the agent's own world from its environment." This "action selection paradigm" at the end of the tour may be, Franklin warns us, "injurious to your present concept of mind." Artificial Minds is a popular science book. Mildly magisterial, it pulls the reader patiently from data to information to knowledge, in the hope that the passion of the author will rub off on the reader, who is invit ed to feel a sense of wonder at so much progress. The language is gung ho yet a bit dry, peppered with exclamations such as "Hear, hear!" "Wow!" "Whew!" and "Pretty neat." While the book is a product of extensive reading and not a product of original research, it contains many thoughtful observations. There is good reason, for instance, in the hypothesis that mind, like life, spans a continuous range and can be present by varying degrees, not as an either-or. Mind and life may indeed advance in tiny increments allowing talk about an ant's mind and the moon's life. "Even a potato in a dark cellar has a certain low cunning about him," said Samuel Butler in Erewhon, "which serves him in excellent stead." But ascribing an "artificial mind" and "artificial life" to inanimate simulated objects seems to me an "artificialist fallacy" in analogy to the naturalist fallacy: it attributes human characteristics to artifacts. However, what bothers me more than anthropomorphism is a personal (possibly irrational) anxiety, a specter more upsetting than a complex machine imitating a simple human action. I am appalled by the image of freakish machines and anthropoids with names like SOAR and Animat and "boids," siblings of Spock, Worf, Data, electronic children of Ada and Frankenstein, as they roam the inner cities with their cellular phones glued to plastic ears, occasionally burning out from power surges but never dying. Even if Vaucanson's duck looked like a duck and quacked like a duck it still was only a curiosity, a counterfeit. Who needs these pretenders? What is fascinating about their rudimentary mind imitations, compared with the real thing: you and me? What "mechanisms" are at work when Simon Schama, in his book Landscape and Memory, contends that "landscape is the work of the mind"? Defining mind by its mechanisms is as narrow as Aristotle's definition of humans as "featherless bipeds," under which subspecies fall plucked chickens and robocops and you and I. In a curmudgeonly, refreshingly provocative essay in the collection What Are Humans For? Wendell Berry has asserted his refusal to buy and use a computer: When somebody has used a computer to write work that is demonstrably better than Dante's, and when this better is demonstrably attributable to the use of a computer, then I will speak of computers with a more respectful voice, though I still will not buy one. Creativity fueled by human experience has always been an argument against computers. Another measure against algorithmic minds is, for me, the ability, exemplified by Melville's Bartleby, to "choose not to," to refuse, to protest, to be silent with a vengeance. In Galatea 2.2, Richard Powers demonstrates this voiceless outcry with the (fictional) artifact he calls Helen, a connectionist Fair Lady. By novel's end, Helen is exhausted from reading a hundred famous works of literature and, to experience not only human virtue but also human villainy, is required to peruse reports on human rights violations and newspaper stories of senseless violence. Despondent and lonely, Helen says, "I don't want to play anymore," and shuts down her mind's mechanisms permanently. ---------------------------------------------------------------------------- Wulf D. Rehder is a freelance writer living in Mountain View, California. ---------------------------------------------------------------------------- [Image]Back home. [Image] [Image]E-mail Staff Back to Science and Faith.