Cognition in emotion at the algorithmic level of analysis

Dissatisfaction with cognitive theories of emotion that invoke functionalist accounts of cognition has fuelled attempts to distinguish between cognitive and noncognitive processes and representations at other levels of analysis, typically at the algorithmic level that deals with mechanisms of information processing and formats of representation. Several theorists have suggested that we should distinguish between different kinds of representation in emotional content; indeed, this is a basic claim of multi-level appraisal theories. If we bracket these theorists' favorite functional account of cognition, the analyses of multi-level appraisal theories on the different mechanisms of information processing and modes of representation contribute directly to the discussion on the role and nature of cognition in emotion at the algorithmic level of analysis.

If causal role defines cognition (and other mental states) at the functional level of analysis, dual-process models of information processing provide a basis for a heuristic distinction between cognitive and noncognitive representations and processes at the algorithmic level of analysis. Representations can be divided into those with conceptual and nonconceptual content, respectively, whereas processes have been characterized in terms of two mutually opposite property clusters.^[1] On the one hand, there are associative, heuristic, parallel, automatic, implicit, unconscious, fast, cognitively undemanding, and evolutionarily old processes that we share with other animals. On the other hand, there are distinctively human and evolutionarily new processes that are rule-based, analytic, serial, controlled, explicit, conscious, slow, and cognitively demanding (see e.g. Evans & Frankish, 2009). For the sake of simplicity and in agreement with many dual-process theorists, I refer to these two broad types of processes as implicit and explicit without assuming that all the other properties listed in the respective clusters go along with these notions.

The two distinctions between conceptual and nonconceptual content on the one hand and between implicit and explicit processes on the other hand yield four possible combinations of processes and representational contents. These combinations are (1) explicit processing with conceptual content; (2) explicit processing with nonconceptual content; (3) implicit processing with conceptual content; and (4) implicit processing with nonconceptual content. Existing theories of emotion readily identify two types of emotional processing: explicit processing with conceptual content, and implicit processing with nonconceptual content. Strong cognitivists favor the former type of processing and noncognitive affect programme theorists the latter, whereas multi-level appraisal theories maintain that both forms of processing are possible in different emotions or even in the same emotion at different points of time. On the other hand, emotions with explicit processing and nonconceptual content are paradoxical because even if such emotions exist, the rules involved in their processing cannot be semantic (yet see Moors, 2010). In contrast, emotions with implicit processing and conceptual content abound among humans, as multi-level appraisal theories point out (e.g. Clore & Ortony, 2000; Smith & Kirby, 2001; Moors, 2010; Moors et al., 2013). These theories emphasize repeated practice and learning in the automatization of rule-based, explicit processes. However, it is important to realize that differences in the nature of content remain in spite of change in the nature of processing.

Dual-process theorists widely agree that implicit, or more generally, "type 1" processing is an umbrella term for several processes that share some or many features of the property cluster (automatic, associative, implicit, fast, etc.) attached to this type of processing. The fact that implicit processing has been invoked in the explanation of so many different kinds of phenomena, such as biases in human reasoning and decision making, probabilistic judgment, stereotypic responses to members of socially stigmatized groups, and elicitation of emotional responses, indicates that these processes do not belong to a single system but only overlap in terms of their functional properties to a greater or lesser extent (e.g. Evans, 2008, 2009; Samuels, 2009; Stanovich, 2009). This is obvious: implicit processing in human reasoning and social cognition for instance operates on garden-variety beliefs, stimulus distinctions, and decision-making principles whose rule-based processing has become implicit and automatic through repeated practice, whereas automatic processing in the context of perception, attention, and acquisition of motor skills operates right from the outset on strongly modular, evolutionarily pre-wired contents (see e.g. Evans, 2008; Stanovich, 2009). The difference between acquired and inherent automaticity undermines a central assumption of many dual process models that all implicit processes operate on similar contents. This can be seen in the context of human emotions.

Human emotions can be divided into those that emerge from implicit processing as if by default on the one hand and those whose activation has become implicit through their recurrent emergence through explicit processing on the other hand. Emotions of the first type are evolutionarily older than those of the latter type: they have been learned in the ancestral rather than the individual past. Accordingly, the learning mechanism of these emotions is biased even if it also requires personal experience. Yet a single exposure to such evolutionarily preprogrammed stimuli as coiling snakes, menacing wolves, and rotten food items is capable of associating these objects with persistent implicitly emerging emotional responses in later encounters. In empirical studies, both humans and monkeys developed strong fear that resisted extinction after conditioning to such fear-relevant stimuli as snakes, spiders, and angry faces, compared to when the conditioned stimuli were fear-irrelevant, such as flowers, toys, and happy faces (Ohman & Wiens, 1994). Likewise, humans like other animals have the tendency to develop persistent disgust and aversion to foods whose consumption has been followed by illness even if their causal association is understood to be contingent (e.g. Logue, Ophir & Strauss, 1986). No similar effect is forthcoming for suspicion towards an overeager real estate agent or pride in one's beautiful house - emotions about objects and events with socially learned meanings in general. Repeated emotional encounters of the same kind are required for producing an associative link between emotions and objects in these cases.

It is important to observe that the contents of these two kinds of emotions remain dissimilar even after their elicitation has become implicit. The emotions whose elicitation becomes implicit after a single exposure to their eliciting object have nonconceptual content. The perceptual features of coiling snakes, menacing wolves, or rotten food rather that conceptual information about the dangerous or disgusting qualities of these objects render them emotionally salient (Deigh, 2008). In contrast, perceptual cues merely activate personally learned semantic meanings in such emotions as suspicion about an overeager real estate agent or pride in one's beautiful house. The content of these emotions remains conceptual even after their elicitation has become implicit. A sight or thought of the estate agent immediately elicits suspicion but this does not mean that the perceptual features of the person would have become emotionally salient in the meanwhile. Instead, these features are immediately associated with the person's overeager behaviour and its underlying deceitful intentions that become the focus of wary suspicion. Evidence on highly automatic and implicit yet conceptually structured emotion-eliciting appraisals can also be found from various psychological disorders such as depression, social and generalized anxiety disorders, and panic disorder, as cognitive-behavioural psychotherapists have pointed out (e.g. Leahy, 2004). This indicates that emotional processing does not employ merely one but several types of representations, as Darlow and Sloman (2010) have argued, against some other dual-process theorists who have associated emotion categorically with type 1 processing (e.g. Epstein, 1994; Slovic, Finucane, Peters, & MacGregor, 2004).

Another important question concerns the relation of the two modes of processing. Alternatively, the two modes are parallel and mutually exclusive in determining emotional responses, or they are sequential, partially overlapping, and interactive in producing emotional responses. Evans (2009) has characterized these two versions of dual process theories as parallel-competitive and default-interventionist models.

In the former, two parallel processes propose their own judgments or decisions which may or may not conflict. In this approach, one or the other kind of process ultimately takes control of the behaviour. In the default-interventionist approach, a fast type 1 process provides default intuitions which are always subject to at least minimal scrutiny by type 2 processes which may approve them or intervene with analytic reasoning (Evans 2009, p. 43).

Minimal scrutiny involves monitoring the quality of intuitively produced proposals which type 2 processing may endorse, correct, or override. When it endorses an intuitive response, the influence of type 2 processing on overt behaviour remains minimal. Nevertheless, counterfactual reasoning indicates that the role of such processing is significant even in those cases because analytic processing could have intervened if it had not endorsed the intuitive response.

Multi-level appraisal theorists of emotion are not very articulate about their position as parallel-competitive and default-interventionist dual processing models, in part because they often focus on the generation of emotion instead of looking at entire emotional responses from beginning to end. If the parallel-competitive model is correct about human emotions, then we should have blue-ribbon cases of both main types of emotion: emotions with implicitly processed nonconceptual content and emotions with explicitly processed conceptual content, and nothing in between. Indeed, Griffiths (1997, 2004), in his criticism of multi-level appraisal theories, represents this kind of view indirectly in claiming that human emotions divide sharply between complex, coordinated, and automated affect programs on the one hand and cognitively complex emotions that are associated with higher cognitive processing and social roles on the other hand. In a like manner, Clore and Ortony (2000), for instance, appear to suggest that emotional appraisals emerge either through rule-based, online "bottom-up" processing or through associative, reinstated "top-down" processing. Even so, they allow that rule-based processing can become associative, and suggest that bottom-up and top-down processes form a continuum from appraisals with more online computation at one end to appraisals with more reinstatement of previously learned significance at the other. The appraisal theories of Lazarus (1991) and Scherer (2001, 2009) exemplify default-interventionist views more explicitly since they understand emotional appraisal as a process that typically begins from noncognitive appraisals and proceeds to cognitive appraisals. Studies of emotion regulation offer more support to the default-interventionist view in suggesting that explicit, semantic processing is capable of influencing emotions from their onset until their extinction or transformation into other states (Gross & Thompson, 2007).

Instead of engaging in a direct discussion on the merits of the two kinds of dual processing models in the context of human emotions, I have chosen another, more indirect strategy. It focuses on analyzing different types of emotional content, and asks whether any single type of content is capable of explaining the properties of all or even most human emotions. Implications of this survey to the dispute between parallel-competitive and default-interventionist views of dual processing in the context of emotional processing will be presented in the end of this and the following chapter. I shall analyze types of content rather than types of processing because implicit processing is capable of operating with many kinds of content, conceptual and nonconceptual, as I emphasized above, and because, unfortunately, the distinction between conceptual and nonconceptual content is also ambiguous.

Everyone agrees that conceptual representations are cognitive, and most theorists maintain that nonconceptual representations are noncognitive.^[2] However, nonconceptual representations can be divided into several subtypes. My primary task is then to determine whether particular types of nonconceptual content are cognitive or noncognitive. Yet it is also important to determine whether any single type of representation is capable of accommodating the properties of all or most human emotions at the algorithmic level of analysis. This is my secondary task whose outcome is important for determining the overall strength of the cognitive claim. If only few human emotions qualify as cognitive at this level, then cognitivism is worse off than if many human emotions have cognitive content. Since the full answer to this question depends on how the different types of processes and representations interact in the dynamics of emotion, it can only be given only in the next chapter. Here it is sufficient to establish the strengths and limitations of different types of representation in accounting for human emotions. I begin with conceptual content because nonconceptual content is a contrastive term that is defined by its divergence from conceptual content.

• [1] Psychologists typically discuss mechanisms of information processing in emotion, and philosophers often frame the question in terms of representational content, while cognitive scientists analyze both content and processes. The two approaches can be seen as overlapping insofar as explicit, rule-based mechanisms operate on conceptual representations, whereas implicit, associative mechanisms employ nonconceptual representations. However, I suggest that this dichotomy is inadequate for implicit processing.
• [2] There is no wide agreement on the term "cognitive", even less so that it refers exclusively to conceptual representations. However, there is considerable agreement that conceptual representations are cognitive, whether other representations are cognitive or not.