Greg Detre
Monday, March 31, 2003
Norman et al.�s basic distinction is between mechanisms of:
cognition � with which we �interpret, understand, reflect upon, and remember things about the world�
affect � with which we �rapidly evaluate events to provide an initial assessment of their valence or overall value with respect to the person�
They discuss the example of a RAID system that becomes �anxious� when one of its disks breaks down, and reacts by starting to take precautions, prioritising extra reliability over performance and alerting a human overseer. This is a paradigmatic example of how emotions can modulate behaviour, e.g. by shifting priorities, employing different resources, generating interrupts or setting goals. We can see how emotions like fear and anger do just this, and indeed in chapter 1 of The Emotion Machine, Minsky discusses how emotions can often act more like filters, e.g. when we fall in love or get angry (when our reaction times can decrease, presumably because higher-level complex processes get ignored).
Norman et al.�s argument for the need for an affective system relies on their distinction between strong and weak methods. In short, strong methods exploit domain specific knowledge and structure, such as when a programmer writes a rescue mechanism to deal with some particular well-understood contingency. Hard-wired knowledge in biological systems comes under the same category. In contrast, weak methods trade efficiency for generality, and are essential whereever the domain or environment is poorly-understood, uncertain, complex or wide in scope, and where some catch-all procedures can be of some benefit. Hill-climbing is given as one such weak method, but we could also point to human-level learning in general as expensive, slow, but hugely flexible. Affective systems are weak methods � when the RAID system appraises its own trustworthiness and feels satisfied or anxious as a result, if its strong methods fail, weak methods can allow the system to fail gracefully and non-disastrously.
Norman quotes the following criticism made of him that �if we were to follow Norman's prescription, our designs would all be usable, but they would also be ugly� in �Emotion & Design: Attractive Things Work Better�[1]. One of his central arguments about the importance of affect in design is based on the interesting, if somewhat simplifying, observation that positive affect enhances creative, breadth-first thinking whereas negative affect focuses cognition, enhancing depth-first processing and minimizing distractions[2]. This provides an example of how weak methods can be seen to play a role in general/creative decision-making, as opposed to just repairs and reactions. Obviously, most interesting systems benefit from employing both strong and weak methods, and the growing thesis (as exemplified in these papers) of many cognitive psychologists is that cognition and affect are similarly inextricable.
The second way in which emotions play a role is in learning, especially with regard to goals and priorities. As Steve Grand[3] puts it, emotions �provide the value by which we measure things. If you don't have a means of measuring how your actions upon the world work out, and whether they're good or bad, you can never learn from them�.
Minsky�s attachment-based learning is one of the most carefully-elaborated accounts of how emotions can play a role in our evaluations of the world, and of our own actions. Normally, when receiving positive or negative reinforcement (whether from the environment or by a stranger) for some action, we learn to alter our methods in some way. However, when praised or rebuked by people that children love or respect, they don�t just feel pleased or dissatisfied, but rather proud or ashamed, and that these feelings accompany the learning of new goals. We need to learn from those around us, because we have evolved to deal with living in complex communities that potentially vary enormously (in climate, culture, language, lifesteyle etc.) by relying less on survival skills that we possess from soon after birth. For this reason, our parents� and elders� views are even more important, not just because we are so dependent on them as children, but because of the need to �download� their wisdom.
He calls this figure of authority, respect or love an �Imprimer� (after Lorenz�s imprinting research with ducks and geese). Such attachment bonds form according to the �speed with which a person responded to an infant and the intensity of the interaction in which he engaged with that infant�, rather than physical care, according to Bowlby. In Tinbergen�s terminology, the presence and praise of an Imprimer has come to serve as an �innate releasing mechanism� for a special system we call the �attachment elevator�, which raises present goals to a higher kind of priority. Our conscience, ideals and values can be seen as the internal evaluations of our actions based on our internal models of what our Imprimers would say, in their absence (see discussion of cognitive levels below). He�s happy to admit that there�s no neuroscientific evidence for Impriming, but argues that there hasn�t been much of a search for it.
All three authors try to further structure the mind hierarchically. This cognitive hierarchy is largely orthogonal to the emotional structure, especially in Sloman�s schema. Sloman and Norman both use three levels:
Reactive
This is the lowest level. Though Sloman and Norman characterise it in slightly different terms, the gist is that it�s fast, lacks the ability to represent or compare counterfactuals or consequences, and plays a valuable but basic stimulus-response role in reaction to the current state. It may be genetically determined or hard-wired, or may allow some sort of modification, learning or short-term habituation. Importantly though, the number of sensors and situations is limited and quickly explodes combinatorial because of its impoverished representational power. Sloman hives off the �alarm� system as separate, presumably in order to highlight how there can be alarms at each level.
Norman�s second (routine level) contains skilled and well-learned, largely �routinised� behaviours, employing considerable processing. In contrast, I think Sloman would probably include most of such learned but low-level behaviours as reactive, unless they involve some counterfactual comparative representation and processing � such representations are more powerful than the �default expectations� that Norman talks about.
A final important difference is Sloman decision to treat the affective components as orthogonal to the cognitive (rather than being anchored within a single level), creating a 3x3 grid. To me, this makes more sense, as illustrated by the popular distinction between primary, secondary (and tertiary) emotions. After all, some of our emotions are global and relatively lacking in semantic content, e.g. contentment, while others are very semantically specific (and presumably higher-level). Norman�s routine-level affective assessment along the three dimensions of positive, negative and (energetic) arousal puzzled me � I couldn�t tell whether this was intended to be biologically plausible, or merely a convenient simplification, but it certainly seems too impoverished to be useful as a model of human affect.
Deliberative
Sloman describes the deliberative level as being able to �represent, analyse, compare, evaluate and react to descriptions of hypothetical future scenarios or possible explanations of previously observed phenomena�, using structured representations with a compositional semantics. The ability to represent counterfactuals, whether imagined, future, past or in different locations, plus the ability to break situations down into objects defines the deliberative layer.
Meta-management (reflective)
However, the deliberative layer is still unable to represent its own operations. This is where the meta-management, reflective or �self-conscious� layer comes in, performing operations upon its own internal representations of its experiences, its physical embodiment, its current behaviour, and the current environment, along with the outputs of planning, reasoning and problem-solving. Usually, the reflective layer has no direct inputs, but relies on influencing behaviour indirectly via the lower levels.
Minsky, in chapter 5, divides this level further into reflective thinking, self-reflective thinking and self-conscious emotions. Reflective thinking focuses on recent ideas, decisions and actions, while the self-reflective level is able to represent its own capabilities, dispositions and goals so that not only can [it] think about what it has done, but about the something that did what was done�. Finally, the highest level is the self-conscious emotional level, which (rather like a super-ego) considers what one�s Imprimers (or one�s internal models of them) would say about past actions.
The discussion of levels shouldn�t lead us to expect complex, autonomous or flexible systems to exhibit correspondingly clean and distinct internal structures. Rather, as Minsky describes, it�s much more likely that different modules will partially represent each other, with the clean flow of control being muddied by loops in some sort of �tangled hierarchy�[4]. Norman�s claim that control (inhibition/activation) only flows downward, while interruptions could flow upwards, seemed particularly impractical. Abstracting such signals into an �alarms� module that spans all three levels, as Sloman does, doesn�t seem like it allows enough influence between levels either. Moreover, if huamn development is anything to go by, it�s likely that different task/domain-specific modules in our brains are capable of higher-level processing and representation than others (especially at certain stages of development).
[1] Norman, D. A. (2002). Emotion and design: Atrractive things work better. Interactions Magazine, ix (4), 36-42. Also available at jnd.org.
[2] He cites:
Ashby, F. G., Isen, A. M., & Turken, A. U. (1999). A neuropsychological theory of positive affect and its influence on cognition. Psychological Review, 106, 529-550.
Isen, A.M. (1993). Positive affect and decision-making. In M. Lewis & J. M. Haviland (Eds.), Handbook of emotions (pp. 261-277). New York: Guilford.
[3] In an interview with Suzy Hansen on salon.com. See also Creation: life and how to make it, Steve Grand.
[4] Hofstadter, Godel, Escher Bach