Parallel or sequential processes? A brief review

The question of how the two systems are related to each other – to which Frankish’s is only one of the many solutions proposed – is one of the biggest issues facing researchers in the field and possibly represents the deepest divide in dual-process theory itself. It raises a number more specific questions: in what order, for instance, do the two systems operate – simultaneously or sequentially? If simultaneously, how do they collaborate? And what happens if they disagree? Strong arguments have been given both for a parallel (i.e. simultaneous) mode of system interaction and for a sequential one. Although these issues do not play a major part in my own theory, I will briefly mention some of the salient points

A parallel relationship would involve a continuous interplay between the two systems, with both competing for attention and acceptance. Sloman (1996: 15) describes experiments in the psychology of reasoning that seem to illustrate this. Here, associative (System 1) and rule-based (System 2) processes yield incompatible results, with the System 1 results persisting however much participants try to ignore them. As Sloman puts it, ‘Both systems seem to try, at least some of the time, to generate a response. The rule-based system can suppress the response of the associative system in the sense that it can overrule it. The associative system however always has its opinion heard and, because of its speed and efficiency, often precedes and thus neutralizes the rule-based response.’ A similar pattern, whereby System 1 makes a sustained and continuous impact on the simultaneously-operating System 2, is also proposed by Smith & DeCoster (2000: 112).

In sequential models of dual-system processing, these questions of overlap and running competition do not apply. As described by Evans (2009: 45-6), the assumption here is that a fast, automatic, System-1 process ‘precedes and shapes’ subsequent conscious, controlled, effortful System-2 reasoning. On this approach, competition between the two systems is resolved by the sequencing pattern used. Thus, System 1’s output – a default mental model produced by heuristic2 processing

2

Differences over the meaning of this term represent another example of confusing terminology current in dual-process work, and perhaps reflect the lack of connection between the fields from which the theory has emerged. According to Evans, the field of social cognition draws a contrast between heuristics – speedy, low-effort cognitive functioning – and the more effortful type required for considered rule application: a distinction, Evans (2009: 36) comments, ‘between a quick and dirty

– is presented for consideration to the analytic processes of System 2. Where this second-stage processing is shallow and careless, Evans explains, the final output is effectively a heuristic response. However, depending on ‘motivation, cognitive ability, instructional set, time available, [my own emphasis] and so on,’ analytic processing may lead to revision and replacement of the initial default model as a result of more effortful reasoning.

Frankish’s conception of a System 2 that is heavily dependent on System 1 for input and motivating factors seems to take the sequential model as far as it can reasonably go while still preserving some connection between the two systems. However, it is only one of the major variations that have evolved within dual-process theory’s developing structure. Other important ones include the role played by memory in the activities of the two systems and, by extension, in their relationship with learning. This dual model of memory is discussed in the next sub-section and, as I hope to show in the following chapter, a further extension links dual-process theory – in particular, System1 – with speech production that is automatically geared to ensuring Addressee relevance.

5.3.4 ‘Separate but interacting’: duality in memory systems

A major contribution here is made by Smith and DeCoster who propose that System- 1 and System-2 processes are linked to separate memory systems that use two

‘fundamentally different’ principles of operation. The first of these – a slow-learning memory system – supports the processing mode associated with System 1, and the other – a rapid-learning memory system – feeds into the operations of System 2.

heuristic form of processing and a slow and careful systematic form [that] could, but need not [author’s emphasis], reflect an architectural distinction.’ As he and other theorists point out, however, System 2 can also make use of heuristics, drawing on a quick and careless thinking style as well as on slow, careful, consideration. A further complication lies in the fact that, in some contexts, ‘heuristics’ refer to rule-based structures, or rules of thumb, that have to be retrieved from memory and evaluated, presumably consciously. An example given by Frankish (ibid: 94) of this deliberate application of a rule is the recognition heuristic, ‘If you have to say which of two items best satisfies some criterion, then choose the one you recognise.’ Within relevance theory, the relevance-theoretic comprehension procedure is seen as an automatic heuristic which operates at a sub-personal (non- conscious) level. Mercier and Sperber have argued that, within their massive modularity framework, higher-order reasoning is also carried out by intuitive argumentative mechanisms which yield inferences about premise-conclusion relations (see Chapter 5.3.5). So their differences from Evans et al. are mainly about the working of System 2. On the parallel vs sequential issue, Sperber favours a ‘pandemonium’ model in which most automatic procedures operate in parallel and compete for processing resources.

Smith & DeCoster (2000: 109) start from the assumption that there is a functional incompatibility between the demands made on humans by existence. On the one hand, they need to record information slowly and incrementally, thus accumulating a large sample of experiences in memory so that expectations and long-term stable knowledge can be based on average, typical environmental properties. This requires a ‘slow-learning memory system which could be termed “schematic” because it matches the typical properties assumed for schema in social and cognitive theories.’ At the same time, however, humans also need to learn new information rapidly, so that a novel experience can be remembered after only one encounter. This calls for a ‘fast-binding’ system that can store episodic records of details of specific

experiences.

As evidence for this hypothesis, Smith & DeCoster point to disassociations observed in amnesia patients between different types of memory-related capability:

specifically, between the capacity to learn and remember general regularities in the environment (e.g. character traits), as distinct from consciously recollecting

particular events. They continue:

A model featuring two separate but interacting memory systems can account for these sorts of evidence. The slow-learning system is a collection of

overlapping systems that are involved in sensory, perceptual, and motor output processes. The operation of these systems generally does not depend on

conscious awareness or attention. These systems are responsible for translating input representations (e.g. visual patterns of letters) to output representations (e.g. word meanings). They also function as memory systems. Learning takes place as the system processes each stimulus and involves small, incremental alterations of representations in ways that facilitate repetition of the same processing… These representations are then used preconsciously to process and interpret new information by categorizing, filling in unobserved details, and the like.

(ibid: 110)

Meanwhile, the fast-learning memory system, which depends on the hippocampus and related brain structures, is responsible for rapidly constructing episodic memories – new representations, resulting from single experiences – that bring together different aspects of an experience or an object in its context.

Smith & DeCoster propose that these differences in the two memory systems directly affect the operations of dual-process theory’s System 1 and System 2. The slow- learning system drives a preconscious, fast-operating ‘associative processing mode’ that functions as a pattern-completion mechanism: here, a stimulus calls to mind the wealth of information and affective associations that have accumulated in the past. The fast-learning system helps support a ‘rule-based processing mode’ which draws on symbolically and culturally transmitted knowledge and rests on human linguistic abilities. Importantly, Smith & DeCoster point out that these abilities, in turn, draw on both underlying memory systems.

As Table 5.2 shows, there are clear overlaps between Smith & DeCoster’s proposal and core aspects of dual-process theory, notably in connection with the issues of processing speed and automaticity. These are particularly highlighted in connection with the rule-based mode, and for the same reason cited in other versions of the theory: the processing bottleneck that System 2 necessarily involves. Here, it is as much the implementation of the ‘rules’ as their retrieval from memory that slows things up: the process of using the retrieved explicit representations to guide

processing is, Smith & DeCoster explain, ‘necessarily sequential and relatively slow (in contrast to the fast, parallel constraint-satisfaction process that can be used with associative knowledge representations). The reason is that only one rule can be explicitly used to guide processing at a time. Rule-based processing is thus more effortful and time-consuming than associative processing’ (ibid: 112).

ASSOCIATIVE PROCESSING RULE-BASED PROCESSING

Draws on associations… Draws on symbolically-represented rules…

… which are structured by similarity and contiguity…

…which are structured by language and logic…

… and learned over many experiences

… and can be learned in just one or a few experiences

Occurs automatically Occurs optionally when capacity and motivation are present

Occurs preconsciously, with awareness of the result of processing

Occurs often with conscious awareness of processing steps

(after Smith & DeCoster, ibid: 111)

Table 5.2: Theoretical properties of associative and rule-based processing modes

There are two important corollaries to this equation of the associative processing mode with fast retrieval speed and the rule-based mode with the reverse. The first can be simply expressed as ‘learn slow, retrieve fast’ vs ‘learn fast, retrieve

slow(er)’. To this, Smith & DeCoster (ibid: 115-6) add an interactive dimension, whereby over time, the results of slow, effortful reasoning can become established in the associative system, and vice versa:

Suppose someone repeatedly uses a step-by-step rule-based process to make an inference or solve a problem – perhaps just counting on one’s fingers to get the answer to 2 + 3. Repeated trials create the conditions for associative learning, so eventually the same answer can be retrieved by pattern-completion from the associative system, rendering the step-by-step procedure superfluous… With

enough practice, therefore, the answer to such a problem just pops into consciousness.

Equally, information can move the other way: people can, perhaps, reflect on their own past accumulated experiences and condense them into a symbolically-

represented rule. For example, frequent and successful fishing trips with a friend may lead someone to realise that the friend always knows where the fish are biting: a conclusion that in turn leads to further considerations, such as how the friend knows. As Smith & DeCoster comment, ‘Turning this knowledge from a mere association built up from repeated experiences into a symbolic representation has several benefits – the knowledge can be used flexibly, applied in other contexts, or it can be communicated to other people.’

What happens when there is a conflict between the outputs of the associative and rule-based systems of this memory-based model? Citing instances where this results from a clash between argument strength (e.g. weak) and appeal of source (e.g. attractive or expert), Smith & DeCoster suggest that the crucial factors that govern the overall response are motivation and cognitive capacity. Given that the rule-based system demands effort and attentional resources, they say, there has to be some motivation – such as a desire for accuracy or argument validity – in order to use it. If this is lacking, the response will generally be determined by the relatively effortless associative system. They also make the further crucial point (ibid: 117) that

‘cognitive capacity’ refers to temporal as well as attentional resources:

Rule-based processing generally takes longer than associative processing… and, because it requires attention, it is more subject to disruption by distraction, interference, and so forth. Thus, responses that are made quickly or when the perceiver is busy or distracted likely will be controlled by the associative system. However, given adequate time and freedom from distraction, rule- based responses (because of their greater subjective validity) may override associative responses.

The second key point that Smith & DeCoster make about the link between the associative system and fast retrieval speed has to do with the huge variety of outputs it can produce. Its pattern-completion mode of operation, operating automatically and preconsciously, has such a wealth of accumulated representations to draw on that even a limited stimulus can activate a wide variety of these, and they may go well

beyond the purely informational. Thus, the sight of a mug activates the intuitive conceptual information that it is used to hold coffee, while the sight of a friend may bring an affective response of warmth and affection.

Another feature of the associative system is that it can use currently available cues to retrieve representations that were stored on past occasions when similar cues were present. ‘Through associative processing, information that has repeatedly been linked to an object in the past is automatically brought to mind whenever we perceive or think about the object again. This information can fill in unobserved details or can even change the way people perceive existing features of an object’ (ibid: 111). In Chapter 6, I consider some of the implications for Addressee-oriented speech of such automatically-retrieved links, whether to objects, situations, or

individuals.