Keywords

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

1 From Psychology to Cognitive Neuroscience, to Modelling, and Back

The middle 1960s have been the cradle of a cognitive revolution (Gardner, 1985). The explicit goal of this movement was to depart from behaviourism, which dominated most of the twentieth century, by posing a new paradigm. Behaviourism, the “science of the black box”, believed that the purpose of psychology was to study and predict behaviour. Little importance, if at all, was put on the underpinnings of the behaviour. Unlike behaviourists, the proponents of the cognitive revolution wanted to figure out how the mind, and ultimately the brain, processes information. First chronicler of this revolution, Gardner (1985) depicted this new scientific landscape as groups of disciplinary researchers attempting to bridge the gaps between their disciplinary expertise (Fig. 1, panel a). His conclusion at the time was that there was no consensual research paradigm, no common set of hypotheses, and no consensual methodology. Things have evolved since then and, even though the utopia of witnessing truly interdisciplinary approaches to the mind is never fully met, strong ties between disciplines are being built.

Fig. 1
figure 4_1_213094_1_En

(a) Connections among the cognitive sciences, after Gardner (1985). Unbroken lines = strong interdisciplinary ties. Broken lines = weak interdisciplinary ties. (b) The triangle of the cognitive neuroscience, adapted from Kosslyn and Koenig (1992). (c) Proposal for an implementation of interdisciplinary dynamics between psychology, neuroscience, and artificial neural networks modelling (Roesch et al., 2007)

Full size image

Proof is the perspective taken by Kosslyn and Koenig (1992) who formalised the cognitive neuroscience triangle (Fig. 1, panel b), in which the three complementary objects of research (behaviours, computations, and brain substrates) are mutually constraining each other. In particular, a general challenge for cognitive neuroscientists, which is considered in the present chapter, is to investigate psychological functioning in a biologically plausible way that can produce computational models – with both the biological and the computational levels constraining psychological models of cognition (Roesch et al. 2007). These constraints can be expressed in the form of the goodness of fit between the models proposed by each discipline (Fig. 1, panel c).

The goal of cognitive neuroscience is very ambitious. It is nothing less than “to map the information-processing structure of the human mind and to discover how this computational organisation is implemented in the physical organization of the brain” (Tooby and Cosmides, 2000). By being as explicit as possible, researchers can account for healthy and pathological functioning and design new research in such a way as to address new predictions. This approach is often referred to as the “boxes-and-arrows” approach , because of the graphical representation researchers use to describe the topology of cognitive systems and the interactions and dynamics between their constituents. Boxes represent the computations that are taking place. They take inputs, and deliver outputs, and interact in fixed manners. Arrows represent the mechanisms of interaction, the dynamics of the information flow (Marr, 1982).

Sander and Koenig (2002) developed structural and functional principles for such a description applied to emotional processing. The constraints they extracted review the basics of emotion literature and contain criteria that all models of emotions should be able to account for. Functional principles encompass criteria related to the evaluation of the stimuli, the expression of the emotion, and the subjective experience of the organism. These are complemented with computational criteria describing the types of input, and output, that would be relevant to emotion models. They also emphasise the need for biological plausibility. The mind being intrinsically tied to the cerebral substrate, biological plausibility constraints posit that, in order to adequately represent cognitive functioning, models should be as close as possible to what is known of the nervous system (Kosslyn and Koenig, 1992). However, some argue that, in a first stage of the modelling process, biological implausibility can be an asset to understand and characterise the problem, examine and evaluate the information flow using a representation that is sometimes easier to grasp (Dror and Gallogly, 1999).

Extending on Sander and Koenig’s approach to the modelling of emotion, one can emphasise a few points and make the following recommendations to both theoretical and applied research groups.

1.1

1.1.1

1.1.1.1
1.1.1.1.1
1.1.1.1.1.1 Recommendation 1 – Formal Description of Endogenous and Exogenous Structures

A model accounts for a specific system (endogenous structure), which only makes sense in a specific context (exogenous structure). Therefore, we argue that models should be embedded in a context as detailed as possible. The main purpose of this principle is to ease the comparison first between models and second with experimental data.

1.1.1.1.1.2 Recommendation 2 – Formal Description of Inputs and Outputs

Inputs can be of several types, coming from the environment of the organism (e.g. visual, tactile) or from internal stimulation (e.g. somatic information). Similarly, outputs can feed forward into deeper processes or backward into the system. This information being at the core of the model, researchers need to provide a detailed description of each arrow they draw.

1.1.1.1.1.3 Recommendation 3 – Formal Description of Transfer Functions

Models describe the interaction of several processes. Each process computes the information it receives and transmits the processed information in an output further. What the process does is not necessarily clear, and the dynamics involved can be very dependent on the interpretation of the researcher. We therefore argue that great effort should be invested into the description of the processes, in relation to the input it receives and the output it transmits.

1.1.1.1.1.4 Recommendation 4 – Formal Description of the Information Flow

The information transmitted may change and evolve. Therefore, we argue that researchers should describe the sequence of actions that occur in the system as precisely as possible. The time course of the information in the model is of critical importance to compare it with experimental data.

1.1.1.1.1.5 Recommendation 5 – Formal Description of State Transitions

At any point in time, a model is in an observable state, expressing a particular mode of behaviour of the system and can only be in that particular state. Consequently, the model could potentially be described by more than one state. We therefore argue that researchers should describe the states the model can be in, as well as the possible transitions that can follow. This recommendation could help the comparison with behavioural and neuroimaging data.

Critically, adopting a cognitive neuroscience approach to emotion may help to distinguish between alternative theories of emotional processing, by providing new constraints that are brain based as much as computation based. Until now, two major classes of psychological theories of emotion have dominated research in cognitive neuroscience of emotion, with an emphasis on brain-based evidence: (1) basic emotions theories and (2) dimensional theories (see also the chapter “Emotion: Concepts and Definitions” by Cowie et al., this volume):

(1) The dominant view of emotion during the last century, and which is probably still the most influential in current emotion research, is represented by the so-called basic emotions models (Ekman, 2003; Izard, 2007). These models convey the notion that there is a finite number of separate emotions, shaped by evolutionary pressure, that can be combined to form blended and more complex emotions. Whereas researchers do not agree on the exact number of such emotions, they all agree that “the various classes of emotion are mediated by separate neural systems” (LeDoux, 1996). Most of the recent cognitive neuroscience research on emotion has attempted to identify specific brain regions implementing these distinct basic emotions (Ekman, 1999) such as fear , disgust , anger , sadness , and happiness . Indeed, a large corpus of data suggests that signals of fear and disgust are processed by distinct neural substrates (see Calder et al., 2001). Functional imaging of the normal human brain (e.g. Phillips et al., 1997) and behavioural investigations of brain-damaged patients (e.g. Calder et al., 2000) revealed a crucial involvement of insula and basal ganglia in processing disgust signals. On the other hand, animal research (e.g. LeDoux, 1996), behavioural studies of brain-damaged patients (e.g. Adolphs et al., 1994), and functional imaging in healthy people (e.g. Morris et al., 1996) suggested that the amygdala is a key structure for responding to fear-related stimuli (see also Öhman and Mineka, 2001). Mineka and Öhman (2002) even proposed that “the amygdala seems to be the central brain area dedicated to the fear module” (see also Öhman and Mineka, 2001). More tentative evidence suggests a similar segregation of processes related to anger, sadness, and happiness, particularly during recognition of facial expression (e.g. Blair et al., 1999). On the basis of neuropsychological dissociations between fear, disgust, and anger, Calder et al. (2001, 2004) encouraged neuropsychologists to adopt the basic emotions framework in order to understand and dissect the emotion system.

(2) From another perspective, all emotions are considered to be represented in a common multidimensional space. For example, Wundt (1905) proposed that the nature of each emotion category is defined by its position within three orthogonal dimensions: pleasantness–unpleasantness, rest–activation, and relaxation–attention. It has been argued that emotional response and stimulus evaluation might primarily be characterised by two dimensions: valence (negative–positive) and intensity (low–high) (see also Anderson and Sobel, 2003; Hamann, 2003). Using functional neuroimaging, Anderson et al. (2003) and Anderson and Sobel (2003) found that amygdala activation correlated with the intensity but not the valence of odours, whereas distinct regions of orbitofrontal cortex were associated with valence independent of intensity. Similarly, Small et al. (2003) dissociated regions responding to taste intensity and taste valence: structures such as the middle insula and amygdala coded for intensity irrespective of valence, whereas other structures such as the orbitofrontal cortex showed valence-specific responses. However, adopting a theoretically based approach, recent results show that four dimensions (valence, potency, arousal, and unpredictability) are needed to satisfactorily represent similarities and differences in the meaning of emotion words, Fontaine et al. (2007) questioning the ability of two-dimensional models to fully represent emotional systems.

However, instead of adopting either the discrete emotions or the dimensional views, some researchers have proposed an alternative approach, by parsing emotions into distinct subcomponents at the process level and determining dynamic interactions between these processes. If Damasio (1994, 1999) distinguished neural systems involved in emotion from those involved in feelings , Panksepp (1998) proposed four primitive systems (seeking, fear, rage, and panic systems) combined with special-purpose socio-emotional systems (for sexual lust, maternal care, and rough-housing play). With an emphasis on psychopathology , Gray (1994) distinguished three types of behaviour (fight, active avoidance, and behavioural inhibition), each mediated by different neural circuits. Similarly, Davidson (1995) proposed differential systems in the two cerebral hemispheres underlying approach-related emotions and withdrawal-related emotions.

These diverse approaches illustrate the lack of a consensual definition of emotion (see Kleinginna and Kleinginna, 1981). With the aim to elaborate on a working definition, Scherer (1984, 2001) integrated different theoretical elements and proposed a multicomponent approach to emotions (see also the chapter “Emotion: Concepts and Definitions” by Cowie et al., this volume). This definition is part of the theoretical framework laid by appraisal theorists of emotion, which represent an attempt to parse the underlying functional mechanisms in terms compatible with brain-based and computation-based evidence.

A key aspect of appraisal is that the specificity and the differentiation of an emotion may depend on a multifactorial evaluation of the meaning and consequences of an event, given the individual’s goals, needs, and values, as well as the current context. To our knowledge, until recently, neither basic emotion theorists nor dimensional theorists have been centrally concerned with the nature of processes that may not only detect relevant events, but also evaluate their consequence and meaning in a context-dependent manner. By contrast, such evaluation is central to componential appraisal theories of emotion (Ellsworth, 1991; Ellsworth and Scherer, 2003; Frijda, 1986; Lazarus, 1999; Roseman and Smith, 2001; Smith and Lazarus, 1993; Scherer, 2001). Such an approach conceptualises the behavioural meaning of an event for the individual (and thus the resulting emotion) on the basis of multiple complementary criteria including novelty, agreeableness, goal conduciveness, coping potential, and norm compatibility (Scherer, 2001). Therefore, a cognitive neuroscience account of appraisal processes in emotion may offer new avenues of investigation and possibly account for results that otherwise remain difficult to explain (Sander et al., 2005). Among the neural networks involved in emotional processing, a few critical structures have been intensely investigated, but their respective roles remain difficult to link with theories of emotion. These structures include the amygdala , ventral striatum, dorsolateral prefrontal cortex , superior temporal sulcus, somatosensory-related cortices, orbitofrontal cortex, medial prefrontal cortex, fusiform gyrus, cerebellum, and anterior cingulate (Adolphs et al., 2002; Adolphs, 2003; Damasio, 1998; Davidson and Irwin, 1999; Pessoa, 2008; Rolls, 1999). As already mentioned, it has been proposed that the insula is particularly involved in processing “disgust” (Calder et al., 2001; Phillips et al., 1997). However, the insula was also found to be activated during the experience of sadness (George et al., 1996), fear conditioning (e.g. Büchel et al., 1999), and processing of fearful faces (Anderson et al., 2003), challenging this “basic emotion” approach. Moreover, Morris et al. (1998) showed that the anterior insula was responsive to increasing intensity of fear in faces, and Phelps et al. (2001) proposed that the insular cortex might “be involved in conveying a cortical representation of fear to the amygdala”. Therefore, it appears that the insula, as a structure, might not be uniquely involved in disgust-related mechanisms. In particular, a specific account of the function of the amygdala, derived from appraisal research, can help to constrain and inform models of emotion. Contrary to the assumption that the amygdala is central to a “fear module” only (Öhman and Mineka, 2001), in accord with a discrete emotion model, some brain imaging studies suggest that this structure contributes to the processing of a much wider range of negative affective stimuli (for a review, see Sander et al., 2003). As the amygdala seems also involved in the processing of positive events, it was suggested that it modulates arousal, independently of the valence of the elicitor (e.g. Anderson and Sobel, 2003) – potentially supporting dimensional theories of emotion. However, it has been shown that equally intense stimuli differentially activate the dorsal amygdala (e.g. Whalen et al., 2001), and that arousal ratings in a patient with an amygdala lesion are impaired for negative, but not positive, emotions (Adolphs 1999). These results seem to contradict the view that the amygdala codes for arousal irrespective of valence. From another perspective, it has been proposed that the computational profile of the human amygdala might meet the core appraisal concept of relevance detection (see Sander et al., 2003), a view that integrates several findings on the amygdala and suggests that it may be central in processing self-relevant information. In general terms, an event is relevant for an organism if it can influence (positively or negatively) the attainment of his or her goals, the satisfaction of his or her needs, the maintenance of his or her own well-being, and the well-being of his or her species. Evaluation of relevance may then elicit the corresponding affective responses and behaviours. These responses may include enhanced sensory analysis and enhanced encoding into memory, as well as autonomic, motor, and cognitive effects. A review of the literature is consistent with this idea that amygdala processes do not respond just to the intrinsic valence or arousal level of a stimulus, but to the subjectively appraised relevance (see Sander et al., 2003, 2005). As an example, one can mention the study by Winston et al. (2005) that demonstrated that the amygdala exhibits an intensity-by-valence interaction in olfactory processing. These authors were able to show that the amygdala responds differentially to high (vs. low)-intensity odours for pleasant and unpleasant smells (as previously reported by Anderson and Sobel, 2003 and Small et al. 2003) but, critically, not for neutral smells. This recent study therefore concluded that the amygdala codes neither for intensity nor for valence per se, but for an integration of these dimensions.

In conclusion, a critical function of the cognitive neuroscience of emotion is to design new experiments specifically testing for critical predictions of these current major psychological theories. In particular, determining whether the amygdala is primarily involved in (1) fear or negative valence, (2), intensity or arousal coding, and (3) relevance detection would, respectively, provide support to (1) discrete emotions theories, (2) dimensional theories, or (3) appraisal theories. Although advanced neuroscience approaches have helped to start testing these current major psychological models, computational approaches will certainly bring to the new field of affective neuroscience critical concepts and methods for testing the respective predictions.

2 Biologically Plausible Artificial Neural Networks to Test Models of Emotion

In addition to neural network modelling on a functional or psychological basis as described above, some researchers have attempted to approach implementable emotion modelling from a neural perspective. The aim of this research is to implement a mechanism with known neural correlates, in order to draw a direct analogy to neural mechanisms. Thus, the ultimate goal is not necessarily to create a network architecture that could have practical applications, but rather to increase our understanding of the workings of the brain by combining the knowledge that we already have in implementable and testable mechanisms. Not only does implementation allow us to grasp the workings of mechanisms that are otherwise too complex to reason through, but it also allows us to assess if non-implemented models are realistic. In this section, we will show how this way of modelling can and has been used to investigate emotional processes.

The architecture of individual neurons or nodes in biologically realistic models is based on the behaviour of neurons or groups of neurons in the brain. A typical example of a single neuron-like node is a leaky integrate and fire neuron that integrates its inputs and fires when a certain threshold is reached. This produces a spiking response , consisting of short bursts of activation. Larger groups of neurons can be represented by a more complex input–output function or architecture, producing a graded response which can be seen as an average of a group of spiking neurons, with overall more gradual changes in the output pattern. Pioneers in the area of biologically realistic neural modelling are Hodgkin and Huxley (1952), who were the first to describe biological neuron data (of the squid giant axon) in mathematical terms in the early 1950s. Their mathematical neuron descriptions were used in many models, which were then again tested experimentally, creating a scientific dialogue between experimental and modelling research. More than half a century later, models of various neural processes that are in one way or another related to emotion, such as conditioning (Grossberg, 1971; Suri and Schultz, 1999; Sutton and Barto, 1981) and attention (Servan-Schreiber et al., 1998; Sperling et al., 2001), have been created, but we are only just beginning to hypothesise about the mechanisms that underlie truly emotional processes.

The following summary of the neural structures involved in affective processes is very short and therefore incomplete and coarse. However, it is necessary to give an introduction to the neural correlates presented in the neural models described later, but it is beyond the scope of this chapter to give a more lengthy overview, hence this brief summary. For a more extensive review, see Lane et al., (2000).

  • Amygdala – As described in the previous section, the amygdala has always been an important structure in neural emotion research and is proven to have a large role in emotional processing, by assigning some kind of value to the perceived objects of the environment.

  • Orbitofrontal cortex (OFC) – This area is known to be active in (emotional) face recognition (Adolphs et al., 2003). It also has a role in conditioning, its activation being related to the value of the expected punishment or reward (Rolls, 2004; Schoenbaum and Roesch, 2005).

  • Anterior cingulate cortex (ACC) – The ACC is involved in decision making and premotor functions. It is important for the production of emotional responses, i.e. arousal (Critchley, 2005), as well as for self-regulation of affect (Phan et al., 2005). It also has a strong role in pain perception (Sewards and Sewards, 2002).

  • Insula – This region is activated during recognition and production of disgust, but (as described above) has also been found to respond to sadness, fear, and reward and has recently also been implicated in the hedonic (feeling) aspect of pain (Sewards and Sewards, 2002).

  • Nucleus accumbens (Nacc) – Part of the ventral striatum, this area is involved in conditioning processes and the anticipation of rewards and punishments (Knutson et al., 2005).

  • Thalamus – This structure relays sensory information to the rest of the brain. Of particular interest in our endeavour, it provides a fast, subcortical pathway to the amygdala.

  • Ventral tegmental area (VTA) – A prediction error signal is generated in this area, which is positive when an unpredicted reward is received and negative when a predicted reward is not received. This signal is governed by the neurotransmitter dopamine.

Based mainly on LeDoux fear conditioning research (LeDoux, 1996), Armony et al. (1995, 1997) developed a range of models specific to fear processing and fear learning, employing cortical and thalamic modulation of the amygdala, which produces behavioural output, reproducing many experimental results in great detail (Fig. 2).

Fig. 2
figure 4_2_213094_1_En

Diagram of the network modelled by Armony et al. (1995, 1997). Brain structures are composed of individual artificial neurons. Patterns of activation are represented with gray shading (black bullets = maximum activation; white bullets = zero activation). Structures are self-inhibitory. Feedforward connections are excitatory, and learning occurs following Hebbian rules

Full size image

This has provided support for the “basic emotions” view of the amygdala as a fear module. However, as described earlier in this chapter, there is evidence that the amygdala also has a role in other emotional processes, which is not at all incompatible with its large role in fear processing. It only means that the role of the amygdala is not restricted to fear.

Another recent example of a simulation of an emotional process is that of Wagar and Thagard (2004), who propose the nucleus accumbens (NAcc) to be a central structure in the integration of cognitive and emotional neural signals for decision making. Modules representing the ventromedial prefrontal cortex (VMPFC), amygdala, hippocampus, and ventral tegmental area (VTA) are also included in this model. VMPFC activation represents a prediction of the outcome of a given response. The amygdala is activated only if this response is emotionally laden, and hippocampal activation simply represents the current context (see Fig. 3 for a schematic outline of connections between these structures in the model). In this model, the NAcc is continuously deactivated by the VTA. This means that more than one input is needed to elicit an output from this structure. So when both the hippocampus and VMPFC are active, which means the VMPFC prediction applies to the current context, a response is elicited. A response can also be elicited when VMPFC and amygdala are simultaneously active, but because these activations tend to be short as they respond to nonpermanent stimuli (unlike the context), there is a very short time frame in which this coactivation can produce a response.

Fig. 3
figure 4_3_213094_1_En

Outline of GAGE model of cognitive-emotional integration (Wagar and Thagard, 2004). VMPFC = ventromedial prefrontal cortex, NAcc = nucleus accumbens, VTA = ventral tegmental area. The VMPFC receives sensory input. The amygdala processes contextual somatic states, and the joint action of the VTA and the hippocampus acts as a gating mechanism over the prefrontal cortex throughput in NAcc neurons

Full size image

Implementation of this model with spiking neurons performing the IOWA gambling task produces NAcc output data that are similar to behavioural data in normal subjects and to data in VMPFC-lesioned patients when the VMPFC module is damaged. This provides a potential mechanism for partly emotion-based decision making, giving a computational account of somatic marker theory (Damasio, 1994) through reciprocal connections between VMPFC and amygdala. In this theory, emotion influences decision making through somatic markers (a kind of emotional memory) associated with the stimulus, stored in the VMPFC. This account is not compatible with basic emotion theories, since amygdala activation represents more than one possible emotional marker. In contrast, neither discrete emotion theories nor an appraisal-based approach is directly contradicted by this model.

Taylor and Fragopanagos (2005) have applied the COrollary Discharge of Attention Movement (CODAM) model (Taylor and Rogers, 2002; Korsten et al., 2006) to emotional phenomena. Originally developed to model consciousness and attention from an engineering control approach, CODAM has proven to be applicable in the modelling of processes with an emotion–attentional overlap, such as the emotional–attentional blink, whereby emotion has an influence on ongoing attentional processing. Figure 4 shows an outline of the CODAM model of the emotional–attentional blink , in which an amygdala module is added to the original CODAM attentional blink model (Fragopanagos et al., 2005) to simulate emotional influences on attentional processing.

Fig. 4
figure 4_4_213094_1_En

CODAM model of the emotional–attentional blink

Full size image

In this model, an attentional control signal is produced by the Inverse Model Controller (IMC) module (analogous to parietal cortical areas), which influences working memory (located in parietal cortex as well) through modulation of the object map (primary and associative cortices). The corollary discharge and monitor module (cingulate cortex) provide error correction to IMC and working memory, partly through comparisons to goals (prefrontal cortex). See the original article for a more detailed description (Fragopanagos et al., 2005). The paradigm simulated with this model is that of Anderson and Phelps (2001), in which the second target in the attentional blink is charged with an emotional load. If this emotional load is sufficient, it cancels the attentional blink that normally renders the second target imperceptible in normal subjects, but not in patients suffering from a lesion of the amygdala. The addition of amygdala modulation in the attentional blink model, and subsequent simulation of amygdala lesions by disabling the amygdala module, produced simulated ERP data that minutely resembled experimental ERP data from both normal subjects and amygdala patients, as well as fMRI data from normal subjects (including amygdala activations). This provides direct support for the theory that the amygdala is the structure creating emotional interference in this task and that it does so through amplification, as in the model.

Based on similar principles, the same authors have also simulated other paradigms where amygdala activation has been connected to emotional influences on attention (Taylor and Fragopanagos, 2005), such as the Pessoa et al. (2002) face/bars divided attention paradigm. The fact that this framework is suitable to model a whole range of emotion–attention paradigms provides support not only for the theory that the amygdala is the core structure in the emotional modulation of attentional processes, but also for the proposed amplification mechanism. Thus, we do gain knowledge not only on the ‘where’ of this emotional process (amygdala) but also on the ‘how’.

The above and numerous other studies have modelled the impact of emotion on attention. However, an important question, of value in getting a better understanding of emotion, is as to if and by how much the reverse occurs. In other words, of the impact of attention on emotion. A series of recent studies (Raymond et al., 2003) have shown that selective attention can influence the emotional value of both selected and ignored items when a complex scene is being attended to. Specifically, ignored items (distracters) were consistently rated less positively in emotional evaluations, following attention selection, relative to (typically) simultaneously presented items (targets). The resulting effect was termed ‘Distracter Devaluation’ (DD), being measured by how much reduction of the value of a distracter occurs when it has just been inhibited in the above search paradigm. Furthermore, a known electrophysiological index of attention selectivity (the so-called N2pc, measuring the degree to which attention has been selectively focused in one hemisphere or another and at about 200–300 ms post-stimulus) was shown to correlate with the magnitude of the observed distracter devaluation.

A neural model was developed by Fragopanagos et al. (2009) to account for these findings by means of a plausible mechanism linking attention processes to emotional evaluations. This mechanism relies on the transformation of attention inhibition of the distracter into a reduction of the value of that distracter (expected to be coded in the orbitofrontal cortex). The model is successful in reproducing the existing behavioural results as well as the observed link between the magnitude of the N2pc (arising in the attention control system) and the magnitude of DD. Moreover, the model proposes a series of testable hypotheses that call for further experimental investigation. In particular it provides a hypothesis as to how inhibitory features arising from prefrontal cortex can be transferred not only as part of the feedback attention control system but simultaneously as part of the rapid manipulation of a value reward system of the emotional limbic system in the brain. As such it begins to cover a different range of time intervals than that associated with the slower conditioned learning of dopamine-assisted predictions of reward much studied in conditioned learning in animals. These models are, again, not compatible with a basic emotion approach, but could fit in a dimensional framework or an appraisal-based approach with the amygdala as relevance detector.

More recent work has focused on the emergence of emotions per se: Korsten et al. (2007) have simulated different representations of the value of self-esteem and connected this valuation to the emergence of anger and, when disrupted, depression. In this model, the module responsible for the comparison between these different self-esteem values is analogous to the cingulate cortex , whereas the values presented to the cingulate cortex module for comparison are thought to be analogous to the prefrontal and orbitofrontal cortices. Expanding and generalising on this initial model, current work focuses on the development of a neural model in which we suggest that emotions arise from comparisons/conflict between three instances of a particular perceived value (e.g. food reward): the ‘actual’ value (the amount of food reward possessed at the moment), corresponding to activation of the amygdala, the ‘expected’ value (the amount of food reward expected to be possessed), activating the orbitofrontal cortex, and the ‘normal’ value (the amount of food reward that we would normally, on average, expect to be possessing), analogous to activation of prefrontal cortex. When a discrepancy between two of these instances is large enough (we are receiving a larger reward than we were expecting), an emotional response emerges (joy, surprise). Through comparisons of these different value instances, it is possible to differentiate between the basic emotions of joy, anger, hope, fear, sadness, disappointment, and relief. Ongoing work focuses on the implementation of this framework.

These three approaches present various detailed theories of the mechanisms underlying emotional processes and processes influenced by emotion. Not every detail of these models has as yet been verified by experimental research, which is why they could provide good guidance as to which aspects of emotional processing could be suitable for neuroscientific research.

3 Conclusion

By providing both a functional account of the processes involved in emotional processing and a description of the underlying cerebral substrate, psychologists and cognitive neuroscientists explain normal and pathological functioning. Several traditions compete, however, and new approaches need to be developed to disentangle their respective predictions. We showed that biologically plausible artificial neural networks can be used in this enterprise by formalising both the structure and the dynamics of the networks involved in the processing of emotional information. Results of such simulations go in line with cognitive theories of emotion, distributing the processing of information over several nodes of complex networks. New predictions can be extracted from these artificial models, and future work will need to involve both experimenters and modellers.