The notion that humans have separate and often conflicting faculties for “cognition” and “emotion” has long fascinated scholars and laymen, as exemplified by Pascal's motto, “The heart has its reasons, which reason does not know,” and by the emotionless (but rational) Star Trek character Mr. Spock.

In cognitive science, the distinction between cognition and emotion is echoed in the idea that segregated brain areas implement cognitive and emotional functions and that there are two independent processing “routes,” one cognitive/controlled and one emotional/automatic, which usually compete (but also occasionally cooperate) to control behavior (Kahneman Reference Kahneman2003a). However useful this “dual-systems” view might be, one would not expect dichotomies inherited from folk psychology – such as cognition versus emotion (or even perception vs. representation vs. action) – map one-to-one to functionally segregated brain areas. Compelling evidence, reviewed in Pessoa's The Cognitive-Emotional Brain (Reference Pessoa2013), challenges the segregation of cognitive and emotional processing in the human brain. The emerging view is that not only cognition interacts with emotion at many levels; but in many respects they are functionally integrated and continuously impact each other's processing.

Studying cognitive-emotional interactions by focusing on advanced cognitive abilities (e.g., how the processing of affectively charged stimuli influences executive control) is important to better understand human cognition. However, this is not the only approach. Cognition and emotion have always been tightly linked; they were linked even before the emergence of sophisticated human abilities. The ancestral reasons for this linkage can be traced back to the demands of situated action control, not of complex cognitive problems; it might be an important requirement of organisms that, as Pavlov (Reference Pavlov1927) has once put it, need to maintain a balance between the internal milieu of the body and the external world.

The basic design of the human brain was largely evident in simpler animals that had to solve the basic problems of survival in situated environments rather than higher cognitive problems. All that mattered for those animals was deploying adaptive behavior to fulfill their needs and motivations – in other words, acting to disclose “desired” or goal states (Cisek Reference Cisek1999; Friston Reference Friston2013; Pezzulo & Castelfranchi Reference Pezzulo and Castelfranchi2009). In turn, pursuing goals requires animals to select what is valuable for them to do (adaptive behavior) or meaningful to process, remember, and predict (perception and cognition). In this perspective, solving the most elementary problems of adaptive behavior requires a close synergy between some (perhaps rudimentary) forms of perception, cognition, emotion, and motivation. Indeed, lower species possess emotional and motivational abilities, here broadly conceptualized as processes that mobilize bodily resources and invigorate actions, control interoceptive signals and autonomic states, bear on the value (and saliency) of stimuli, goals, and information gain, modulate selection processes, and deal with ecologically meaningful events such as rewards and punishments or their expectations (Panksepp Reference Panksepp2011; Rolls Reference Rolls, Fellous and Arbib2005).

Importantly, in order to be useful for adaptive behavior, the contribution of emotion and motivation has to be integrated in a timely manner within the organism's action-perception cycle. What is currently valuable is highly contextual and governed by numerous factors that entail perception and cognition, including metabolic costs, affordances, and the available repertoire of choices (e.g., for a gazelle, finding food versus escaping from a lion). Considering all of these factors represents a serious challenge for architectures that segregate perception, action, and utility (Cisek & Pastor-Bernier Reference Cisek and Pastor-Bernier2014; Lepora & Pezzulo, in press; Verschure et al. Reference Verschure, Pennartz and Pezzulo2014).

The evolutionary perspective pursued here implies that the very fabric of cognition – from its ancestral origins to the most modern and sophisticated skills – is inextricably linked to utility, adaptivity, and meaningfulness; and in turn these should imbue cognitive processing at large, as testified by the ubiquitous presence of value- and outcome-related signals in the brain (Vickery et al. Reference Vickery, Chun and Lee2011). Furthermore, this perspective suggests that cognitive-emotional interactions can be better understood within a systems-level framework that explains how cognition and emotion resolve basic adaptive problems and how elaborations of these basic processes underpin cognitive abilities.

Active inference offers one such framework (Friston Reference Friston2010). Active inference considers that organisms act to fulfil prior beliefs or expectations that encode the (evolutionary) values of their states (e.g., having access to food). In this framework, the brain is a statistical organ that learns the structure of the (external and internal) world and the causes of perception (and interoception) in the form of hierarchical models. Such models support both perception (predictive coding) and action (active inference) in a seamless fashion by minimizing free energy or, more simply, prediction error.

Perception corresponds to (Bayesian) inference that tests cognitive hypotheses – encoded at higher hierarchical levels – against sensory evidence. The higher levels generate predictions in a top-down manner, and the discrepancy (prediction error) between descending predictions and sensations compete to update perceptual hypotheses (i.e., select the hypothesis that minimizes prediction error). Crucially, prediction errors can also be minimized through action. Here, active inference enters the scene and adds two novel ingredients to the picture (Friston et al. Reference Friston, Daunizeau and Kiebel2009). First, top-down predictions can act as higher level goals for the organism (e.g., being satiated); and, second, goals steer a cascading flow of exteroceptive and proprioceptive predictions that are fulfilled by peripheral (motor and autonomic) reflexes – until the goal state is achieved.

Up to now we have focused on the minimization of exteroceptive and proprioceptive prediction errors. However, the same principles apply to the regulation of interoceptive and bodily information that links to the autonomic system to sympathetic and parasympathetic systems. This interoceptive inference has been linked to emotion and self-awareness (Friston et al. Reference Friston, Stephan, Montague and Dolan2014; Seth Reference Seth2013; Seth et al. Reference Seth, Suzuki and Critchley2012). In this view, the internal world is controlled by autonomic reflexes that fulfill descending interoceptive predictions – where descending predictions provide homoeostatic setpoints. These interoceptive predictions are one aspect of multimodal predictions from high-level beliefs about our embodied self (literally our “gut feelings”). In this formulation, interoceptive information does not cause self awareness, or vice versa: There is a circular causality in which hierarchical representations enslave autonomic reflexes, whereas interoceptive prediction errors inform hierarchical representations. Emotion can therefore be regarded as a necessary attribute of any representation that engenders interoceptive predictions.

To characterize the adaptive function of emotions and feelings (Damasio & Carvalho Reference Damasio and Carvalho2013) – and their integration within the action-perception cycle – we consider how adaptive behavioral control rests on the functional integration of exteroceptive, proprioceptive, and interoceptive signals. These have complementary roles: Interoceptive prediction errors inform the current motivational need (or drive) of an animal, in terms of a discrepancy between optimal homeostatic levels (e.g., hunger in terms of low glucose), whereas exteroceptive and proprioceptive prediction errors specify allostatic goals in terms of the sensory states the animal has to solicit by acting (e.g., the sensations of eating). Suppressing the exteroceptive prediction errors (that report the fact that I am not currently eating) can be resolved by acting (eating), which in turn suppresses interoceptive prediction errors engendered by an empty stomach.

That simple scenario exemplifies how hierarchical inference supports motivated behavior, while laying the foundations of emotional-cognitive integration. In turn, this integration spans other domains of perception, cognition, and affect. For example, interoceptive and exteroceptive streams can be integrated to support a “cognitive-emotional” inference. In this embodied predictive coding perspective (Pezzulo Reference Pezzulo2013), the most plausible causes of events are inferred based on both exteroceptive (what I see) and interoceptive (how do I feel) cues. An affectively charged event, such as the presence of a predator can be recognized and categorized from both its perceptual characteristics and the fear it instills in us – with a form of perception that is not “pure” but affectively biased (Barrett & Bar Reference Barrett and Bar2009). This in turn induces a circular causality; where fear is both a cause and a consequence of (predator) perception. The temporal span of this process offers various opportunities to impinge on other processes such as memory and planning.

Note that in embodied predictive coding, some “cognitive” processes are off-loaded to the body, as bodily states (e.g., high arousal state or heart rate) become part and parcel of the inference and can influence it (Garfinkel et al. Reference Garfinkel, Minati, Gray, Seth, Dolan and Critchley2014). The potential adaptive value of this body-based mechanism is apparent in dangerous situations, where high arousal can prioritize cognitive processing. Note also that patients with congenital insensitivity to pain have problems in recognizing potentially harmful scenarios.

Interoceptive signals can report motivational urges, too, so that the mismatch between an internal need (e.g., hunger) and sensory stimuli (e.g., no food) can modulate the importance of visual signals (Montague & King-Casas Reference Montague and King-Casas2007) and focus attention (Mysore & Knudsen Reference Mysore and Knudsen2011). In active inference, these dynamics depend on a precision-weighting mechanism that underpins all hierarchical inference. This mechanism prioritizes top-down ascending prediction errors depending on their precision (inverse variance), and – as a consequence – it regulates various competitive processes (Desimone & Duncan Reference Desimone and Duncan1995), both within levels (e.g., between perceptual hypotheses) and across levels (e.g., giving prominence to prior beliefs over sensory evidence). In addition, precision dynamics regulate the balance between goal achievement and belief revision, because expectations having high precisions are immune to revision by prediction errors and thus act as immutable goals that steer action.

Precision-weighting has been linked to neuromodulation and the control of attention (in the perceptual domain) and affordance (in action selection) (Feldman & Friston Reference Feldman and Friston2010). Precision-weighting is a key source of cognitive-emotional modulations, too. Supposing that valuable goals or needs are reported by exteroceptive or interoceptive signals with high precision, they focus perceptual processing on events that have behavioral significance. Furthermore, they induce a cascade of top-down predictions and subsequent action (Friston et al. Reference Friston, Shiner, FitzGerald, Galea, Adams, Brown, Dolan, Moran, Stephan and Bestmann2012). More broadly, precision-weighting is a flexible mechanism through which an organism's goals and needs prioritize processing dimensions for cognition and emotion: which events should be attended to and predicted (attention and planning) and what are their outcomes and affective consequences (emotion); which expectations have value and should be fulfilled (goal achievement); which prediction errors should be monitored and corrected (cognitive control); and what should be learned from them (memory). By the same mechanism, conflicts between “cognition” and “emotion” can be created in psychological experiments; for example, by using affectively charged but task-irrelevant stimuli as “distractors.”

We have described how precision dynamics prioritize active inference depending on affective or motivational value. In a similar way, at the longer timescale of learning, neuronal coding resources can be prioritized to capture the most behaviorally significant events in the long-range statistics of interoceptive and exteroceptive signals. This phenomenon might produce an “affective tuning” of neural representations at multiple levels. For example, (Machens et al. Reference Machens, Gollisch, Kolesnikova and Herz2005) report that auditory neurons in the grasshopper are particularly sensitive to a behaviorally relevant stimuli – mating signals – rather than representing the unbiased distribution of natural sounds. At deeper hierarchical levels, neuronal populations might encode expectations of states that have value for the organism (goal states) with higher resolution. This has cascading consequences on perception and action because – in active inference – precise expectations at higher hierarchical levels enslave behavior. Furthermore – pursuing an embodied view of cognition – the same hierarchies supporting action-perception loops might be reused for more advanced cognitive abilities such as planning, mindreading, and executive function (Barsalou Reference Barsalou2008; Jeannerod Reference Jeannerod2006; Pezzulo Reference Pezzulo2012; Reference Pezzulo2014; Pezzulo et al. Reference Pezzulo, Barsalou, Cangelosi, Fischer, McRae and Spivey2011; Reference Pezzulo, Barsalou, Cangelosi, Fischer, McRae and Spivey2013; Reference Pezzulo, van der Meer, Lansink and Pennartz2014). This suggests that the same “affective tuning” of neuronal hierarchies naturally biases higher cognitive abilities, too.

In sum, various active inference mechanisms reviewed here – embodied predictive coding, precision dynamics, and the affective tuning of neural representations – offer a rich substrate for cognitive-emotional interactions. At the same time, they are prone to vulnerabilities and misregulations, which might produce psychopathological states such as anxiety disorders and psychotic symptoms (Adams et al. Reference Adams, Stephan, Brown, Frith and Friston2013; Friston et al. Reference Friston, Stephan, Montague and Dolan2014). Some forms of psychopathology might depend on (precision-mediated) misregulations at the cognitive-emotional interface. For example, eating disorders could arise from the failure to handle conflicting information at the level of interoceptive signals (e.g., hunger) versus body schema and (false) belief (e.g., seeing oneself as fat). In active inference, this deficit is not purely representational but determines how subjects act (and think); for example, anorexics plan their actions as if they had an enlarged body (Keizer et al. Reference Keizer, Smeets, Dijkerman, Uzunbajakau, van Elburg and Postma2013). Conceptualizing psychopathological states in terms of active inference may shed light on this intriguing domain, where cognitive-emotional interactions are clearly evident in both a clinical and neuropsychological sense.