Ocorreu um erro neste gadget

sexta-feira, março 16, 2007

Por que escolher este blog?

Sem muito tempo para blogar hoje, estou com Leonardo e Raphael - que fez 4 anos anteontem. Assim, vai aqui uma resenha da Nature que achei interessante, especialmente para aqueles que têm uma mulher viciada em bingo como eu.


Does human creativity stem from a process that turns arbitrary ideas into goals like food and sex?

BOOK REVIEWEDWhy Choose This Book? How We Make Decisions

by Read Montague

Dutton: 2006. 331 pp. $24.95, £15.99

Choose your own reward

R. BAMBER/ALAMY

Neat trick: activities such as gambling and card games can supplant the basic rewards of food or sex.

Why don't our brains get as hot as the processors in our personal computers? And what does that tell us about biological computation, the nature of choice, the value of value signals, and the power of ideas? These questions may seem rather disparate, but Read Montague's provocative and accessible treatment of them in Why Choose This Book? displays a deep and unexpected unity. The story goes something like this. Biological computation is constrained to be efficient: to confer close to maximal computational power for close to minimal energetic expenditure. The outward sign of this efficiency is said to be the mere warmth of the human brain compared with the searing (wasted) heat of those computer processors. To be efficient, biological computations need to be equipped, so the argument goes, with some kind of measure of their own value, in relation to the in-built goals of maintaining life and reproductive success. Such measures allow the system to expend energy only on those computations that matter most. How such a measure might work remains problematic, but once it is in place, general principles of thrifty processing, such as the slow use of power, compression of data, conservation of wiring, and frugal use of bandwidth and communication, are all recruited to the mix.

But it is the goals and value signals that play the lead role in Montague's story. He introduces us to the guiding principle that will link efficiency to choice and to the power (and pathologies) of ideas. That principle, familiar enough in cognitive scientific circles but here tweaked and nuanced in novel and potentially transformative ways, goes by the unpromising name of 'reinforcement learning'. In reinforcement learning, goal states are approached by sensitivity to signals that predict rewards (the attainment of goals). But the system is not simply hardwired to regard only some fixed set of signals as reward signals; it obtains flexibility by learning associations between experienced signals and temporally removed (but consistently associated) rewards. Past experience of what signal leads to what reward is thus combined with present feedback (what's here now, and what is it worth?) to generate choices that (ideally) maximize total future reward.

Moving all this along is the 'reward-prediction error signal', which carries information about how well the actual rewards tally with the predicted rewards. When the actual reward exceeds the predicted one, it makes sense to upgrade the stored value of the states that predicted the unexpectedly greater reward. In the brain, dopamine neurons provide at least one means of mechanistically encoding just such a reward-prediction error signal. Bursts of dopaminergic activity result when the reward exceeds the predicted reward; pauses in activity mean that the reward falls short of the predicted reward; and unchanged activity means the reward was as expected.

But what counts as a reward anyway? The most obvious rewards are the basic biological achievements of life maintenance (such as the ingestion of a tasty and nourishing morsel) and reproduction (or rather its precursor, sexual intercourse). Montague is motivated, however, by a strong desire to unravel the mechanistic underpinnings of what he describes as a uniquely human 'superpower': the capacity to make choices that seem to value biologically arbitrary objects, achievements and actions. Examples of such biologically arbitrary goal states mentioned in the text include solving Fermat's last theorem and committing group suicide in the belief that a spaceship hidden in a comet's tail will then take you to 'the next level'. What makes all this possible, in Montague's model, is the capacity of ideas themselves to act as reward signals, hijacking the prediction-error systems implemented by dopamine neurons in the brain. When this happens, the dopamine outputs start to act as error signals that encourage the rest of the brain to learn and to make decisions in ways that increase the chances of acquiring some biologically arbitrary reward.

Given the potentially biologically catastrophic consequences of such re-tooling of mere thoughts as rewards, Montague suggests that powerful filtering processes control what gets into the reward slot. But such processes can be fooled — in ways that the book describes in compelling and often sinister detail — by damage, by drug abuse, and perhaps even by some forms of advertising and branding (brands are just cues that predict rewards). Montague's proposal is that biologically arbitrary goals can somehow plug into a kind of 'special status reward socket', and thus become a basic, primary reward, like food or sex. He does not claim that these ideas become associated, either directly or indirectly, with food or sex; rather, they plug directly into the 'socket' normally occupied only by the most basic high-status rewards. If we humans have indeed learnt such a powerful trick, it is no surprise that it fuels so much that is both good (creative and expansive) and ill (pathological and restrictive) in our species. Montague begins by laying out this possibility, then follows it deep into the fascinating territories of creative thought, addiction, obsessive–compulsive disorder, Parkinson's disease, and then on to the psychosocial realms of trust and regret.

Despite its attractions, there are some important mechanistic gaps in the story, as Montague acknowledges. For example, it isn't clear why or how one idea might win out over another in the bid to occupy a high-status reward socket, or how the occupation itself is accomplished. Nor is it really clear when such occupation should be deemed pathological rather than creative. I was also left wondering whether the basic idea of each symbol and each computation carrying its own value 'tag' — the difference, Montague argues, between standard computational models using 'meaningless symbols' and the hyper-efficient, value-rich computations said to be characteristic of biological nervous systems — is sufficiently clear and workable. Exactly how do these computation-value or symbol-value pairs work, and how do they transform mere symbol processing into meaning? Do they compose? Two computations whose individuals values are low might together constitute a complex computation whose value to the organism is high, but Montague suggests no way of systematically predicting such combined values from the values assigned to the parts.

Perhaps I am missing something, but it repeatedly struck me that Montague's overall vision is both rather more radical, and rather less mechanistically clear, than his book suggests. The prospective reader should be aware that the story on offer actually departs quite a long way from the basic computational theory of the mind. It builds in value and computation right down to the cellular level, and (more generally) systematically blurs the usual distinctions between life, mind and information processing. This blurring is evident, for example, in the puzzling idea that each individual neuron, in the quest for efficient interneural communication, might need to contain up to 100 million 'dynamic models' of other neurons and neuronal subsystems.

These are not really complaints, however. The book spans several seldom-bridged worlds, from neuroscience to psychiatry, economics and social psychology, and does so with wit, precision and elegance. It succeeds in many of its goals. Above all, it left me feeling I had actually learnt something about myself: a thinking, feeling, choosing, yet painfully vulnerable chemically modulated learning machine.

  1. Andy Clark is at the School of Philosophy, Psychology and Language Sciences, University of Edinburgh, Edinburgh EH8 9JX, UK.

Nenhum comentário: