Historical Pitfalls and New Directions in the Neuroscience of Emotion

The brain basis of emotion: A brief history

The neuroscientific study of emotion began in the 19^th century, when psychology was transformed from mental philosophy into a full-fledged scientific discipline. Scientists searched for the physical basis of fear, anger, sadness, among other mental categories, using the tools of physiology and neurology. During this era, the science of emotion was guided by a Linnaean-type taxonomy of emotion categories, a small set of which were each cast as independent mental organs with a unique neural circuit and/or pattern of physiological correlates (i.e., a physical fingerprint) waiting to be discovered.¹ This family of approaches (Barrett, 2016; Gross & Barrett, 2011), which we refer to as the classical view of emotions (Barrett, 2017a,c),² is an example of faculty psychology – the belief that the human mind is structured as a set of mental abilities, each associated with a unique state caused by its own, unique biological system.³ In the 19^th century, faculties were the categories that described what the mind is (as a formal ontology), what the mind does (as a set of psychological processes) and how the mind is caused (i.e., a functional architecture of biological processes that create the psychological processes and corresponding mental phenomena). Scientists formed grand theories as they battled over whether emotions were constituted as biological changes in body or in brain (e.g., Cannon, 1915). A compromise was struck—emotions were assigned to live in the ancient parts of the brain that control the body, dubbed the “limbic system”, a.k.a., our inner beast, whereas cognitions were assigned to the cortex (mistakenly called the neocortex), like a crown designed for us by evolution (e.g., Papez, 1937). Eventually, these ideas became known as the triune brain concept (Maclean, 1990).⁴

In the late 19^th and early 20^th centuries, studies accumulated in search of the biobehavioral basis of each emotion category, and scientists discovered that they were unable to identify any specific neural circuit or ensemble of autonomic nervous system changes (or facial movements, for that matter) that consistently distinguished the instances of one emotion category from another (for a review, see Gendron & Barrett, 2009). Some studies reported no biological or behavioral differences between categories, but more common was variation in distinctiveness – for example, the physical pattern for anger that distinguished it from sadness, fear and so on in one study was different from the pattern observed in other studies. Variation was the norm (Barrett, 2006a, 2013, 2017c). To deal with this variability, scientists recast emotions as functional states to be studied solely by manipulating their observable causes and measuring their observable effects.⁵ Anger, sadness, fear and so on were ill-defined (or defiantly not defined at all, e.g.,⁶). That is the beauty of functionalism – a phenomenon can be studied as a set of causes and effects without ever operationalizing the phenomenon itself in scientific terms as a set of physical changes.

In the first half of the 20^th century, scientific studies guided by functionalism continued to produce a plethora of highly variable, contextual findings. This quickly gave way to behaviorism, when the mind disappeared as a topic of scientific study altogether, and neuroscience focused on the biology of behavior. Correspondingly, emotions were no longer considered mental phenomena but were instead recast as states of the nervous system that caused specific behaviors. The neural circuitry for freezing or fleeing became the circuitry for fear. The neural circuitry for fighting became the circuitry for anger. And so on. Emotion categories, like all psychological categories, were ontologically reduced to behaviors. Experience and other mental phenomena became epiphenomenal to emotion, if they were considered at all.

Psychology emerged from behaviorism in the 1950s and 60s and the mind was reinstated as a topic of scientific inquiry. However, it was faculty psychology all over again, this time with computers (rather than organs of the body) as the driving metaphor. Emotions were recast as computations (either metaphorically or literally). For example, an “affect program” (Tomkins, 1962, 1963) for anger was said to be an inherited central organizing mechanism that produced a state of anger (Ekman, 1972 and later Ekman & Cordaro, 2011). A “primary process” system for fear was said to produce a state of fear (Panksepp, 1982 and later Panksepp, 1998; Panksepp & Watt, 2011). An evaluation of novelty (referred to as an “appraisal” of novelty) was said to produce a state of novelty (Scherer, 1984 and later Scherer, 2009). Once again, an emotion was equated with the process that caused it. Emotion categories remained ontologically reduced to behaviors, not because experience was irrelevant, but in the name of evolution; non-human animals clearly have emotions, the argument went, but they do not necessarily have emotional experiences, and scientists wanted a species-general explanation of emotion to generalize their findings from rats to humans.⁷

In the early 1990s, the cognitive revolution gave birth to the field of cognitive neuroscience, with affective neuroscience following soon after. Faculty psychology and its taxonomy of mental categories still reigned (Lindquist & Barrett, 2012) and the goal (again) was to carve the brain into a set of independently functioning mental organs: one set of neurons dedicated to each mental category. Accordingly, distinct sets of neurons were thought to cause distinct processes that produced distinct states of the same name. Just as memories were thought to be produced by memory processes localized to memory circuits, attention was thought to be produced by attentional processes localized to attention circuits, and sensations produced by sensory processes localized to sensory circuits, so too emotions were thought to be caused by emotion processes localized to emotion circuits. The hypothesis that each emotion category is caused by its own dedicated neural circuit is a hallmark of the classical emotion view, according to a recent review, which states that “agreed-upon gold standard is the presence of neurons dedicated to the emotion’s activation” (Tracy & Randles, 2011, p. 398). In fact, only some versions of the classical view entertain this hypothesis (e.g., the most recent version of Ekman’s basic emotion theory no longer explicitly hypothesizes that each emotion category has a dedicated neural circuit; see Ekman & Cordaro, 2011).⁸ The triune brain concept was retained in the classical view, sometimes implicitly, for example, with the prefrontal cortex (as the putative home of cognition) downregulating the amygdala (as the putative home of emotion).

The current state of things

History has repeated itself. As neuroscientific studies of emotion accumulated over the last twenty years, it seemed at first as if the neurons responsible for each anger, sadness, fear, and so on could each be localized to specific brain regions (e.g., Murphy et al., 2003). It quickly became clear, however, that scientists were facing the same challenges as a century earlier: studies that were purposely designed to isolate the specific neural basis of a mental category and distinguish it from other categories were consistently unable to do so; although an individual study, on its own, might appear to support the neural specificity of a given emotion category, meta-analytic evidence does not support this hypothesis (e.g., Kober et al., 2008; Lindquist et al., 2012).

Take, for example, links between fear and the amygdala. Much of what we know about the amygdala’s role in fear in humans comes from studying Patient S.M., who lost both her amygdalae to Urbach-Wiethe disease (for a review of findings and relevant references, see an excellent review by Feinstein et al., 2016).⁹ S.M. has difficulty experiencing fear in many normative circumstances (e.g. horror movies, haunted houses), but she experiences intense fear during experiments where she is asked to breathe air with higher concentrations of CO₂. She mounts a normal skin conductance response to an unexpectedly loud sound, but her brain seems unable to use arousal as a learning cue in milder laboratory situations (e.g. what used to be called ‘fear learning’ paradigms) and she therefore has difficulty learning from prior errors (e.g. she does not mount an anticipatory skin conductance response to aversive stimuli, during the Iowa Gambling Task, and she does not show loss aversion when gambling). Interestingly, however, there is evidence from S.M.’s everyday life that she is capable of aversive learning (i.e., ‘fear learning’) in certain circumstances. For example, S.M. is averse to breaking the law for fear of getting in trouble. She also spontaneously reports feeling worried. She ‘learns fear’ in the real world in certain circumstances, evidenced by the fact that she avoids seeking medical treatment and visiting the dentist because of pain she experienced on a previous occasion. Yet, S.M. also shows evidence of failure to “learn fear” in everyday life; for example, she returned to a park where she was attacked at knifepoint the previous day (Feinstein et al., 2011). Therefore, an intact amygdala does not appear to be necessary to organize defensive responses to perceived threats in all contexts. This pattern of findings is consistent with the existence of degenerate circuitry for fear (Barrett, 2017c): degeneracy exists when structurally distinct mechanisms perform the same function (Edelman & Gally, 2001). For example, animal species that have been engineered to knockout selected genes have some number of individuals (up to 30%) who continue to show the phenotype despite the absence of the selected genes (see Edelman & Gally, 2001). Another example of degeneracy for fear can be found in the case of monozygotic twins, both of whom suffered amygdala damage from Urbach-Wiethe disease, but only one of whom shows problems with processing fear-inducing stimuli (Becker et al., 2012).

The hypothesis that there may be degenerate circuitry supporting the emergence of fear is also supported by the study of non-human animals. For example, macaques monkey with amygdala lesions (in the central nucleus) freeze less when exposed to a potential threat (called the human intruder paradigm; Kalin, 2017).¹⁰ Other studies indicate, however, that macaques with neonatal amygdala ablations show blunted responses to all evocative stimuli (not specific to threat; Bliss-Moreau et al., 2011) and are able to learn about new threats via associative learning (i.e., ‘fear learning’; see Bliss-Moreau, Moadab, & Amaral, 2016; Bliss-Moreau, Lavenex, & Amaral (under review) so that their socio-affective function appears to be relatively normal by adulthood (Bliss-Moreau et al., 2017). In addition, there is some evidence that, in certain conditions, rodents without an intact amygdala can show normal responses in situations involving threat (e.g., McGaugh, 2016).

Reviews of neuroscience research within a variety of measurement domains generally do not support the hypothesis that an individual emotion category can be consistently and specifically localized to a specific swath of brain tissue. The circuitry for a given emotion category has yet to be consistently and specifically localized to individual neurons (e.g., see Guillory & Bujarski, 2014 for a review of intracranial recording research on humans; see Barrett et al., 2007 for a review of electrical stimulation research on non-human animals), to a brain region (for a recent meta-analysis, see Lindquist et al., 2012), or to a brain network (for a review, see Barrett & Satpute, 2013; for an example of an individual experiment, see Touroutolgou et al., 2015). A growing number of brain imaging studies have successfully distinguished emotion categories from one another using multivariate patterns of activity distributed across the brain (i.e., pattern classification), but again, when it comes to the observed pattern for any single emotion category, variation is the norm. The diagnostic patterns observed for anger, sadness, fear, disgust and so on, are highly variable from study to study (e.g., compare patterns reported in Kassam et al., 2013; Kragel & LaBar, 2015; Saarimaki et al., 2015 and the patterns derived from meta-analyses as reported in Wager et al., 2015).

Taken together, these findings call the classical view of emotion into doubt.¹¹ In response, functionalism has again been invoked to sustain the viability of the classical view (e.g., Adolphs, 2017; Anderson & Adolphs, 2014). Functionalism introduces a host of scientific problems, however, the most serious being that it is rooted in psychological essentialism that renders the classical view non-falsifiable (Barrett, 2017b). Psychological essentialism (Medin & Ortony, 1989) allows people to define a psychological phenomenon by its causes and consequences while positing a hypothetical state and its hypothetical causal mechanism. People continue to believe in the existence of the state and its causal mechanisms, even when they remain unmeasurable and unspecified. Ekman’s metaphorical affect program for fear, Panksepp’s hypothetical FEAR system, and Adolph’s proposed ‘central fear state’ are all examples of psychological essentialism.

The problems of essentialism are also on display when scientists interpret pattern classifiers as neural “signatures” (Kassam et al., 2013), “fingerprints” (Saarimaki et al., 2015), and “biomarkers” (Kragel & LaBar, 2015) of emotion categories. Such words imply that the pattern for each emotion category is equivalent to the brain state for that category -- something like a neural essence – such that successfully diagnosed instances of the category must contain its features that are both unique to that category and present in every instance.¹² Such interpretations represents a misunderstanding of pattern classification, however (Clark-Polner et al., 2016). The derived patterns for each emotion category need not be observed in every (or even any) single instance of that category that is successfully classified (for an explanation and mathematical simulation, see Clark-Polner et al., 2016).¹³

Just as brain imaging studies find substantially different patterns of brain activity distinguish different emotion categories from one another across studies, studies that successfully distinguish between emotion categories using ANS measures report different patterns from one study to the next, even when those studies use the same stimuli and methods, and sample participants from the same population (e.g., compare findings from Kragel & LaBar, 2013 with Stephens, Christie, & Friedman, 2010). Empirical support for the classical view requires a unique pattern of physiological correlates (i.e., a physical fingerprint) that is consistent and specific for each emotion category. Instead, the body of published findings gives evidence of variable patterns for each emotion category across different contexts.¹⁴ This is not evidence that some studies support the classical view of emotion while others fail to. Instead, it is evidence that the physical correlates of emotion, be they neural or autonomic, are highly variable within an emotion category from instance to instance (and the physiological changes associated with one emotion category in one study can be similar to the changes associated with a different emotion category in a different study). It is evidence that, when it comes to the biobehavioral correlates of any emotion category, variation is the norm. This degree of variability demands a mechanistic explanation that goes beyond explaining it away as epiphenomenal to emotion (i.e., it is not merely due to emotion regulation, display rules, insufficient laboratory methods, etc.).

Lesion and optogenetic studies of freezing, fighting, and fleeing in non-human animals are often interpreted as providing strong support for the unique neural basis of different emotion categories. A closer inspection reveals variable findings, consistent with variation observed in humans. Again, turning to studies of behaviors that have been defined as “fear,” various circuits acting in parallel or collaboratively support learning when to perform these behaviors. Studies have documented many (circuits) to one (behavior) mappings while others find one (circuit) to many (behavior) mappings (Herry & Johansen, 2014; Tovote et al., 2015). Still other studies find that the amygdala, or specific parts (e.g. the basolateral nuclei) are not necessary for the expression of previously learned aversive behaviors (i.e. the way that learning is expressed depends on the context and the available options for behavior; McGaugh, 2016).

Furthermore, consider that animals perform many different behaviors in response to threat, each of which is supported by its own specific circuitry. For example, rats avoid the location of uncertain threat when they are free to move around, such as in a testing chamber with several arms (e.g., Vazdarjanova & McGaugh, 1998); when they are not free to move around, they freeze. When threat is certain and they cannot escape, rats respond with defensive aggression – they kick bedding towards the threatening object (e.g., Reynolds & Berridge, 2008) or jump on it and bite (e.g., Blanchard et al., 1989). And the circuitry that supports defensive aggression towards a cat (a predator) is distinct from that supporting defensive aggression towards another rat (a dominant intruder in the cage; Motta et al., 2009). Freezing, fleeing, and defensive aggression towards various predators are all responses to potential danger, so which corresponding circuit is the real fear circuit? Is this evidence of many fear circuits? If so, then what is the scientific value of the category “fear,” other than its obvious utility for communicating research findings to people who do not study emotion? How do we know that defensive aggression is an instance of fear and not an instance of anger? Obviously, differential behavioral responses are not random – actions are contextually situated (Bolles & Fanselow, 1980; Fanselow, 1994). Therefore, context is important for understanding the variable biobehavioral patterns within any emotion category, as well as the potential similarities across categories.

A more serious issue is whether neuroscience studies of behavior in non-human animals can ever be used to test the hypothesis of specific neural circuits for emotion categories in the first place. The neural circuitry for freezing only reveals something about “fear” when you equate emotion with behavior and stipulate, at the outset, that freezing occurs during a state of fear (and only during fear). Equating freezing with fear is an example of mentalizing on the part of the scientist – it is inferring a mental state that is not directly visible. Thus, interpreting neural evidence about circuitry for behavior as evidence for the brain basis of emotion rests on the scientific justification and validity of mental inference (Barrett, 2017c; LeDoux, 2013). The role of mental inference in interpreting neuroscience research on non-human animals is rarely discussed explicitly, and as a consequence, many scientists who work in that domain routinely fail to recognize the role that their own (often implicit) mental inferences automatically play in interpreting their data. They refer to circuits as controlling fear when in fact they are studying context-dependent behaviors that may not have a one-to-one correspondence with a category of fear. Animal models that allow for probing the brain basis of behavior may be extremely useful for providing hypotheses about the neurobiology of human emotions (even if there is a many-to-many mapping of actions to emotion categories; e.g., Shackman & Fox, 2016).

One way forward is to do away with emotion words and just study the variable instances themselves. One lesson of behaviorism, however, is that a science of instances is not viable. Induction and generalizability depend on the ability to form categories. The problem is not in forming a category – it is in essentializing the category. A category is a group of objects or instances that are treated as similar for some purpose. To essentialize is to believe that these similarities are fixed and real in nature in a perceiver-independent (or ontologically objective) way rather than created by a human mind. Essentialism is the belief that category members (like instances of fear) all share a deep, unchanging feature (or set of features) -- an essence -- giving the category its identity. Any attempt to localize a mental category to a single set of dedicated of neurons, in any form, rests on essentialism. From the standpoint of the classical view, the way to deal with the tremendous variation observed thus far in the neuroscientific study of emotion is to create more fine-grained taxonomies, attempting to bring nature under control and make it easier to identify the unique neural essence of each emotion category. Yet it is unclear whether this approach has thus far produced a generalizable and generative science of emotion.

A multi-level constructionist approach to the study of emotion

Another approach is to follow the data rather than subjugate it. The weight of evidence across different neuroscientific domains suggests the need for a contextually sensitive, constructionist approach to understanding the neurobiological basis of emotion. In the science of emotion, past constructionist approaches largely ignored biology: social construction focused on how emotions are influenced by social roles and values and psychological construction focused on how emotions emerge from more basic psychological processes related to making meaning of affective feelings. But a new breed of constructionist approaches is working to integrate social and psychological construction with neural construction (research that focuses on how experience-driven brain development wires a brain that can create emotions) (for examples, see Barrett & Russell, 2015; Cunningham et al., 2013; LeDoux, 2015; LeDoux & Pine, 2017; LeDoux & Brown, 2017; Lindquist, 2013; Satpute & Lindquist, under review; Satpute et al., 2013). The theory of constructed emotion is an example of this approach (see Barrett, 2017c, d), and we use it here as a guiding example for the future of neuroscientific investigations of emotion to give a wide-ranging biologically plausible account of how emotions, as mental events, are constructed.¹⁵

A recognition that definitions of emotion are stipulated, not discovered

In philosophy, a stipulation is definition by fiat. The history of emotion research reveals that emotions are treated as discoveries, but in fact they are prescriptions. Scientists began with a mental framework for emotion (using categories from mental philosophy) which then dictates the sorts of questions asked, experiments designed, and interpretations offered (and even what counted as data in the first place). It is only this stipulation -- a mental inference linking observed actions (e.g., freezing) to certain function or goals (e.g., for fear) -- that allows scientists to claim that the circuitry for the actions is evidence for emotion circuits. What we learn (or fail to learn) about emotion in any experiment is determined by how we define emotions in the first place. Rather than starting with emotion categories and searching for their physical basis in the brain and the rest of the nervous system, the theory of constructed emotion begins with what is known about the structural and functional organization of nervous systems more generally and asks how they create emotions (i.e. a reverse engineering approach). Ultimately, this approach is revealing that emotional events are constructed by domain general processes, meaning that the category of emotion is not a natural kind (Barrett, 2009), just as emotion categories themselves are not natural kinds (Barrett, 2006a; Barrett et al., 2007).

Allostasis, interoception and affect are at the core of the brain

A brain is one big structure composed of neurons that pass information back and forth to one another. This structure is bathed in a chemical system (supported by “helper” glial cells) that continually modifies the ease and speed with which neurons share information. Information flow is also managed, in part, by inhibitory neurons whose activity fluctuates dynamically in time (e.g., Deneve & Machens, 2016; Sillito, 1975) rather than by a static modularity that is anatomically built into its architecture. Ongoing fluctuations in information flow allow this single physical structure to realize an astoundingly large number of spatiotemporal patterns as neurons dynamically coordinate their activity into different coherent states (i.e., the brain is a complex system). Part of the brain’s complexity comes from its ability to generate novel representations by combining (“remembering”) bits and pieces of past experience; in short, the brain is a complex, information gaining system (Bullmore & Sporns, 2009; Tononi et al., 2008). This more dynamic systems-based view of the brain makes a compelling a case for moving away from faculty psychology’s search for independent mental organs in the brain and thereby relinquishing essentialism (because an instance of perception, emotion, or any mental event, is better thought of as a brain state than as one brain region passing perceptual information to another, which computes emotion, after which it passes information to yet another brain regions, which launches behavior (see Barrett, 2009)).

Furthermore, brains did not evolve for rationality, happiness, or accurate perception. Brains evolved, in part, to efficiently ensure resources for physiological systems within an animal’s body (i.e., its internal milieu) so that an animal can grow, survive, and reproduce (Sterling & Laughlin, 2015). This balancing act is called allostasis (Sterling, 2012). The brain networks that are most important for implementing allostasis (called the “default mode” and “salience” networks) also represent the sensory consequences of allostasis (i.e., interoception; Barrett & Simmons, 2015; Kleckner et al., 2017; for related discussions, see (Pezzulo, Rigoli, & Friston, 2015; Seth, 2013; Seth, Suzuki, & Critchley, 2012).

The two interconnected intrinsic brain networks that regulate allostasis and represent interoception are also at the core of many other psychological phenomena, including memory, decision making, theory of mind, attention, and a host of others (Kleckner et al., 2017), and form the backbone for neural communication in the brain (van den Heuval & Sporns, 2013). Such findings prompt the intriguing hypothesis that, whatever else a brain is doing – thinking, feeling, perceiving, preparing for action – it is also implementing allostasis in the service of behavior -- regulating your autonomic nervous system, your immune system, and your endocrine system -- as well as representing interoception.

Interoception is made available to consciousness as lower dimensional feelings of affect (Barrett, 2017; Barrett & Bliss-Moreau, 2009). As a consequence, the properties of affective feelings—valence and arousal (Russell & Barrett, 1999; Kuppens, Tuerlinckx, Russell, & Barrett, 2013)—are basic features of consciousness (Damasio, 1996; Dreyfus & Thompson, 2007; Edelman & Tononi, 2000; James, 1890, 1905; Searle, 1992, 2004; Wundt, 1897) that, importantly, are not unique to instances of emotion. We hypothesize that emotions (as opposed to cognitions) are constructed when interoceptive sensations are intense or when the change in affect is large and foregrounded in awareness. This may help explain why, in mammals, the brain regions that are responsible for establishing and maintaining allostasis (“limbic” regions such as the amygdala, ventral striatum, insula, orbitofrontal cortex, anterior cingulate cortex, and medial prefrontal cortex) are usually assumed to contain the circuits for emotion – they are important for regulating metabolism and energy expenditures associated with intense affect; for a discussion, see Kleckner et al. 2017). In fact, many of these limbic regions happen to be some of the most highly connected regions in the cortex (Van den Heuvel & Sporns, 2011), and they exchange information with midbrain, brainstem, and spinal cord nuclei that coordinate autonomic, immune, and endocrine systems with one another, and with the systems that control skeletomotor movements (Levinthal & Strick, 2012).

This constructionist approach shifts the focus in neuroscientific investigations of emotion in two related ways. First, an identity relationship between an emotion state and its physical causes is no longer assumed. Second, there is the potential to unite cognitive, perceptual, social and affective domains of neuroscience into a unified framework that places metabolism and energy regulation at the core of all mental activity (Barrett, 2017c, d; Barrett & Simmons, 2015; Kleckner et al., 2017). As a consequence, traditional issues surrounding reverse inference (Poldrack, 2006) become a feature instead of a problem (for a similar view, see Fox et al., in press; Okon-Singer et al., in press). All mental events appear to be constructed within a domain-general functional architecture of the brain (Barrett & Satpute, 2013; Mattar et al., 2015). What we learn about emotion may be generalizable to understanding the mechanisms for other mental phenomena (and vice versa).

A focus on functionally integrated brain systems

As a network, the brain has large-scale topographical and dynamic properties. Both structurally and functionally, it can be understood as a set of dynamic communities or subnetworks (e.g., Mattar et al., 2015; Ciric, Nomi, Uddin, & Satpute, in press) which, confusingly, are themselves usually referred to as “networks” (and so we follow that tradition in this paper). Brain networks have the property of homeostasis, meaning that there is a shifting population of neurons that maintain a network over time (e.g., Marder & Taylor, 2011). Accordingly, a fruitful approach may be to understand emotional events as dynamic configurations (i.e., “recipes) of networks and their shifting neural constituents (Barrett, 2009; Lindquist & Barrett, 2012). The brain’s network architecture is dynamic, not static, but it is reasonably conserved across different brain states (see Cloe et al., 2014; Mattar et al., 2015; Ciric, Nomi, Uddin, & Satpute, in press). A key direction of future research, then, is understanding the spatiotemporal dynamics of network integration that creates the brain states that implement individual emotional events.

A focus on meaning-making

Emotions are not reactions to the world. We hypothesize that they are constructions of the world (or more specifically, they are your brain’s construction of your bodily sensations and movements in the immediate context). They are a brain’s explanation for bodily sensations in relation to the surrounding situation. Therefore, the biology of meaning making should figure prominently in the neuroscientific understanding of what emotions are, how they are caused and how they work. The theory of constructed emotion, as a biopsychosocial approach, explicitly includes the neural architecture that implements meaning making (Barrett, 2009, 2012, 2017c, d). For example, neuroimaging studies of emotion consistently reveal that the default mode network is engaged as people experience emotion and simulate behaviors that are labeled as emotional (Lindquist et al., 2012); these findings are instructive, because they suggest that the same circuitry that is necessary for running a mental model of the world (Buckner, 2012), and that is responsible for implementing semantic knowledge (Binder & Desai, 2011; also see Satpute, Badre, Ochsner, 2014; Satpute & Lindquist, under review), for mentalizing (Denny et al., 2012; Spunt, Satpute, & Lieberman, 2011; Spunt & Adolphs, 2015; Satpute & Lieberman, 2006; Amodio & Frith, 2006), and for creating multimodal concepts (Fernandino et al., 2016), including emotion concepts (Peelen, Atkinson, & Vuilleumier, 2010; Skerry & Saxe, 2014; Wilson-Mendenhall et al, 2011), is also important to constructing instances of emotion. Of course, more research is necessary to identify the mechanisms of meaning making in the emergence of emotions. Nonetheless, the neuroscientific study of emotion has something to teach us about the biology of meaning-making more generally.

A focus on prediction

Meaning making within the brain is a predictive activity (Lochmann & Deneve, 2011). In contrast to traditional stimulus-response “feedforward” frameworks, recent predictive coding (a.k.a. active inference, belief propagation) models suggest that the brain functions as Bayesian filters for incoming sensory input, guiding action and constructing perception (e.g., Barrett & Simmons, 2015; Chanes & Barrett, 2016; Clark, 2013; Deneve & Jardri, 2016; Friston, 2010; Hohwy, 2013; Rao & Ballard, 1999; Seth, 2013; Seth, Suzuki, & Critchley, 2011).¹⁶ Past experiences are reconstructed as partial neural patterns that serve as prediction signals (also known as “top-down” or “feedback” signals, and more recently as “forward” models) to continuously anticipate events in the sensory environment, in part, by planning for motor and visceromotor changes.¹⁷ Without past experience as a guide, the brain cannot transform flashes of light into sights, chemicals into smells, and variable air pressure into music, or interoception into emotions, and the result is experiential blindness.¹⁸ Prediction signals are embodied, whole brain representations that anticipate (1) upcoming sensory events both inside the body and out as well as (2) the best action to deal with the impending sensory events. From this perspective, unanticipated information from the world (prediction error) functions as feedback that can be learned to modify predictions (also known as ‘bottom-up’ or, confusingly, ‘feedforward’ signals). Error signals track the difference between the predicted sensations and those that are incoming from the sensory world (including the body’s internal milieu). Once these errors are minimized, predictions also serve as inferences about the causes of sensory events and plans for how to move the body (or not) to deal with them (Lochmann and Deneve, 2011; Hohwy, 2013). By modulating ongoing motor and visceromotor actions to deal with upcoming sensory events, a brain infers their likely causes. Neuroscience evidence supports predictive coding accounts for every exteroceptive sensory system (see Chanes & Barrett, 2016 for examples), and the same predictive dynamics may also be true for sensations arising from inside the body (Barrett & Simmons, 2015; Kleckner et al., 2017).

From this perspective, emotions are not organized reactions to the world. They are guesses about what to do next, rooted in prior experience, and the sensory consequences of those guesses (as well as the sights, sounds, smells and other experiences of the world). Emotions may be the Bayesian filters, predictions, or active inferences (what we have referred to as embodied concepts)—the representations that typically dominate as intrinsic brain activity. The neural representation of emotion, then, may be created from memory, rather than merely associated with or triggered by memory (Barrett, 2017c, d); this hypothesis is consistent with meta-analytic evidence that the neural basis of emotions routinely involves traditional memory-related networks such as the default mode network (Kober et al., 2008; Lindquist et al., 2012; Wager et al., 2015; Wilson-Mendenhall et al., 2015). An abundance of behavioral work, and a handful of neuroimaging studies (e.g. Ochsner et al., 2009; Satpute et al., 2016), have emphasized the role of top-down knowledge in emotion construction (see reviews by Lindquist, Gendron, & Satpute, 2016; Lindquist, Satpute, & Gendron 2015; Satpute & Lindquist, under review). But few, if any, studies properly test a predictive coding perspective on emotion (in part because experimental studies are designed as a sequence of independent stimulus-response trials that force the brain into a mode where prediction error dominates; Barrett, 2017c). Current models posit that predictions and prediction errors are interwoven throughout subcortical and cortical brain regions and organized by lamina (Barrett, 2017d; Barrett & Simmons, 2015; Chanes & Barrett, 2016), which scientists can now observe using high-resolution 7T scanning (Goense, Merkle, & Logothetis, 2012; Satpute et al., 2013). Moving beyond standard neuroimaging studies, a predictive coding inspired model of emotion requires a dynamic, high-resolution, and distributed approach.

A focus on variability, rather than typologies, biomarkers, affect programs, and central states

Because emotion categories do not appear to have a person-independent set of markers (in the autonomic nervous system, in behaviors, or in facial movements), scientific investigations should be explicitly designed to capture the variety within an emotion category, ideally using multiple measures in multiple contexts. This is an example of population thinking, which is one of the conceptual innovations found in Darwin’s On the Origin of Species (for discussion, see Barrett, 2013, 2017c).¹⁹

Consistent with this approach, a growing number of neuroimaging studies now reveal that, when it comes to emotion, variety is the norm and attempt to study this variety in a situated, contextually variable way (e.g., Oosterwijk et al., 2016; Wilson-Mendenhall et al., 2015). A given emotion is a category of variable instances that are tuned to the situation at hand. As situations vary so do the instances of an emotion category. People smile when sad, cry when angry, and scream when happy. A person can tremble, jump, freeze, scream, gasp, hide, attack, and even laugh all in the face of fear. Just instances of the same emotion category can vary across contexts, they can also show variability within the same contexts. This form of variation -- degeneracy (Edelman & Gally, 2001) -- has almost been ignored brain imaging experiments of emotion (although degenerate circuitry for behaviors can be found in some of the non-human animal work, as described above). Degeneracy is ubiquitous in biology (for examples, see Edelman & Gally, 2001; Marder & Taylor, 2011; Tononi, Sporns, & Edelman, 1999), and natural selection favors systems with degeneracy because they are high in complexity and robust to damage (Whitacre & Bender, 2010). A context change reveals a hidden reservoir of variation for selection processes to act on immediately, without waiting for reproduction and mutation. When it comes to the brain, this translates into preserved emotional function in the face of damage and disease (for examples, see Barrett, 2017c). Thus, future neuroscience investigations of emotion might focus more on the capacity for dissimilar representations (e.g., different sets of neurons) to give rise to instances of the same mental category (e.g., anger) in the same context (i.e., many-to-one mappings of structure to function).

Understanding the neural basis of fear (or any emotion category) will benefit from models that are designed to explain all forms of variability. Such variation is unlikely to rely on a core “fear” circuit (i.e. neural essence) with minor fluctuations or modifications, nor be captured by a handful of narrow “fear” categories. The collection of instances referred to as “fear” is more likely to rely on a distributed architecture, no individual piece of which is essential across instances. Nor across individuals (given the likelihood of degeneracy). Future work would benefit from taking an idiographic (individual and instance dependent) approach is preferable to a nomothetic approach.

Conclusion

In all areas of natural science -- physics, chemistry, and biology -- progress has involved a shift away from the essentialism of classical views towards a more dynamic, contextual and constructionist approach to the physical world. Neuroscience, with advances in neuroimaging methods and analysis, are following a similar path, achieving astounding discoveries, some of which transform our basic understanding of how brain creates mind. From this perspective, emotions should be studied as dynamic, highly variable whole brain constructions of what bodily sensations mean in the context of the immediate environment. The mechanisms that implement these constructions are not specific to the domain of emotion but operate across cognitions, perceptions, and action, so that the scientific knowledge of one domain is of value for understanding the others. The next generation of neuroscience investigations, with a focus on whole-brain network dynamics along with time-varying, multivariate analytic approaches, is well-positioned to make landmark discoveries about the neural basis of emotion.

Highlights.

• A dynamic, systems-based neuroscience view of emotion is proposed.

• An emotion word refers to a population of highly variable, situated instances.

• Emotions are not reactions to the world, but are predictions corrected by sensory inputs.

• Prediction signals are concepts that categorize and explain the causes of sensations.

• Affect is a property of consciousness and is not synonymous with emotion.