Predictive coding for the actions and emotions of others and its deficits in autism spectrum disorders

Traditionally, the neural basis of social perception has been studied by showing participants brief examples of the actions or emotions of others presented in randomized order to prevent participants from anticipating what others do and feel. This approach is optimal to isolate the importance of information flow from lower to higher cortical areas. The degree to which feedback connections and Bayesian hierarchical predictive coding contribute to how mammals process more complex social stimuli has been less explored, and will be the focus of this review. We illustrate paradigms that start to capture how participants predict the actions and emotions of others under more ecological conditions, and discuss the brain activity measurement methods suitable to reveal the importance of feedback connections in these predictions. Together, these efforts draw a richer picture of social cognition in which predictive coding and feedback connections play significant roles. We further discuss how the notion of predicting coding is influencing how we think of autism spectrum disorder.


Origins of predictive coding in neuroscience
Traditional models of how the brain processes sensory input, and social stimuli in particular, have emphasized bottom-up processing, i.e. computations associated with a direction of information flow from lower to higher brain regions in the cortical hierarchy (see Felleman and Van Essen, 1991 for a definition of lower and higher cortical regions).Since Hubel and Wiesel (1962), many propose that when we see someone grasping an orange, retinal input is sent to thalamic center/surround cells, several of which converge onto primary visual cortex (V1) cells to create line detectors, several of which then converge to create complex cells that respond to more complex features (e.g.T or L shapes), and so on, to ultimately converge onto temporal lobe neurons that respond specifically to a hand in a particular configuration.In parallel, thalamic neurons also feed onto neurons responding to motion in a particular direction in V1, V2 and middle temporal (MT) cortex, the output of which then integrates into neurons in the temporal lobe that respond selectively to the combinations of movements that characterize reaching and grasping, then finally onto parietal and premotor mirror neurons where the visual bottom-up information converge onto neurons with matching motor properties (Giese and Poggio, 2003).The computations of social perception are done exclusively through the feedforward flow of information creating ever more complex feature detectors (Felleman and Van Essen, 1991;Giese and Poggio, 2003), and perception is seen as an 'outside in' process, driven relatively directly by the stimulus itself.
An alternative view can be found in Von Helmholtz (1867): he suggested human perception is rather the inference of what could have caused the input: if we see a two-dimensional image of overlapping objects, we unconsciously infer what arrangement of real objects could account for such a 2D image, and start to perceive the 3D objects with depth that we inferred.In the 1980s, this view permeated computational models of perception: McClelland and Rumelhart (1981) explained why letters are recognized better and more rapidly when in words than in isolation, by assuming a bidirectional information flow.When seeing the letters composing a word, the letters trigger activity in word nodes containing such letters in natural language.Once word-level nodes are activated, top-down information flow excites letter nodes that are included in the word and inhibits those not included in the word.Perception becomes a dynamic combination of bottom-up and top-down information flow that can explain why and how letters are recognized better and faster when in words.
Based on the observation that feedback connections within the cortical hierarchy are often as abundant as the feedforward connections (Felleman and Van Essen, 1991;Rao and Ballard, 1999) then proposed "a model of visual processing in which feedback connections from a higher-to a lower-order visual cortical area carry predictions of lower-level neural activities, whereas the feedforward connections carry the residual errors between the predictions and the actual lower-level activities".Since, evidence that the processing in early visual cortices depends on feedback from higher cortical regions has started to accumulate, initially through cooling (e.g.Hupé et al., 1998), than through optogenetic inhibition of higher visual cortices(e.g.Kirchberger et al., 2021).In visual neuroscience, predictive coding is becoming main-stream.The degree to which feedback and predictive coding apply to the processing of complex social stimuli outside the visual cortices, including parietal association cortices and frontal premotor systems, has been less explored, and will be the focus of this review.
Over the past years, hierarchical predictive coding has increasingly been merged with notions of Bayesian inference (Aitchison and Lengyel, 2017).Bayesian inference is a normative mathematical framework describing how to update beliefs based on data (Box 1).It states that beliefs about the state of the world (mathematically formalized as a parameter θ) after seeing data, i.e. posterior beliefs, are a distribution of probabilities associated with each possible value of θ, and should be calculated as the product of the prior probability of each possible state with an updating factor reflecting how much better that value of θ predicts the data than the other possible values (Keysers et al., 2020).This formal Bayesian frameworks often predicts behavior well (Fiser et al., 2010), and hierarchical predictive coding architectures can generate computations that approximate Bayesian inferences, which is why these two levels of description have been increasingly integrated to provide complementary levels of understanding of how the brain interacts with the world (Clark, 2013).Embracing Bayesian inferences emphasizes that the brain should entertain multiple beliefs simultaneously with their respective probabilities reflecting their relative plausibility (Pouget et al., 2013), and that prior beliefs and uncertainty should influence the impact of data on posterior beliefs: if we have strong beliefs about θ to start with, the posterior beliefs will be more influenced by the prior; if our prior knowledge about the parameter is weak, the posterior belief will be more influenced by the data.In the extreme, if our prior belief is that a certain state of θ is impossible, even if we see data highly likely under that state, this probability is multiplied by zero (the prior probability), and the posterior probability remains zero.
Most recently, predictive coding has been supplemented with the notion of active inference and the free energy principle to accommodate that animals are not passive perceivers of the world -our motor system shapes our sensory input.If our hand searches for our keys amongst the contents of our pockets, we move our fingers in ways that we hope to generate diagnostic sensory input -action and perception become a unified process.Although the details of active inference and free energy go beyond this review (see Clark, 2013;Friston, 2010 for details), two aspects are relevant to predictions in social cognition: integrating action and perception provides traction on how brains develop internal models that associate motor programs with expectations about upcoming sensory consequences; and predicts that brains actively explore the most diagnostic parts of the environment.
Predictive coding may be particularly important and complex when participants actually interact in real time with others.So called "secondperson" accounts in social neuroscience (Redcay and Schilbach, 2019) focus on the brain mechanisms underlying direct social interactions, and have argued for the importance of creating meaningful interactions in the lab to unravel the mechanisms underlying the bidirectional influence across interaction partners (e.g.Han et al., 2019).These approaches are contrasted against "third-person" approaches in which participants observe social stimuli they cannot directly influence.In a second-person neuroscience context, active inference accounts are particularly fruitful in explaining how individuals mutually influence each other's predictive models (Lehmann et al., 2022): in conversations, we often choose action by forming beliefs about which action will best resolve uncertainty and ambiguity.This decision-making process requires a deep exploration of potential actions and their predicted outcomes -a process rendered complex due to the need to infer others' mental states, involving nested, reciprocal inferences (de Bruin and Michael, 2021).For example, if one person believes another is thinking about their thoughts, this recursive process influences social interaction and action selection.Furthermore, social interactions are inherently dynamic and carry important information based on the timing of actions, such as, for example, the duration of conversational pauses.Such second-person experiments may be particularly suited to probing and understanding social atypicalities, i.e. difficulties in social interactions seen in autism spectrum disorders, that are sometimes most pronounced during real time social interactions.
Finally, predictive coding is also at the core of dominant learning theories.The study of dopaminergic neurons in the monkey revealed that these neurons respond to rewards in ways that do not reflect the amount of reward received on a particular trial, but the reward prediction error, i.e. the difference between how much reward could be predicted based on the past reward history and the actual reward received (Schultz, 2002).The presence of such reward prediction errors demonstrates that the brain actually predicts upcoming rewards, and subtracts them from incoming sensory signals about actual rewards to calculate prediction errors, and has brought predictive coding to the center of how animals learn.In particular, building on the formal classical conditioning theory of Rescorla and Wagner (Rescorla and Wagner, 1972), this has led to the notion that animals choose between alternative actions based on the expected reward value (EV) of each action, and that they learn to predict that expected value on trial t+1 (EV t+1 ) by updating their expectation based on a learning rate times the prediction error (PE) of the past trial.The prediction error is simply the experienced outcome of trial t (O t ) minus the expected value on trial t, so that EV t+1 =LR*PE t =LR*(O t -EV t ) (Box 1).
While predictive coding and Bayesian inference have been tested quite extensively in early sensory cortices, in what follows, we will summarize an increasing evidence base to suggest that predictive coding may also apply to how we perceive social stimuli, including specifically the actions (in Section 2) and perhaps also the emotions (in Section 3) of others, and how we learn about the value of alternatives in the world based on the rewards experienced by others (in Section 3).In Section 4, we will also address an emergent trend to look at disorders of social perception in terms of altered predictive coding and Bayesian inference using the example of autism.

Theoretical proposals
Since the discovery of mirror neurons in the premotor (Gallese et al., 1996;Keysers et al., 2003;Kohler et al., 2002;Umiltaà et al., 2001) and later parietal cortex (Fogassi et al., 2005) of macaques, the neural circuitry involved in the perception of hand actions of other individuals has received substantial attention.Viewing the actions of others contrasted against appropriate visual control stimuli reliably triggers activity in a distributed and well mapped network of brain regions.In macaques, it recruits neurons in the temporal lobe (including cortices lining the superior temporal sulcus, STS and the inferior temporal lobe, IT), the rostral posterior parietal lobe (areas PFG and AIP) and premotor cortex (area F5) (see Rizzolatti and Sinigaglia, 2016, for a review).In human, it includes a high level visual cluster at the border between the lateral occipital lobe and the mid temporal gyrus, multiple clusters in the supramarginal gyrus and BA2, and several premotor cortices in BA6 and BA44 (Caspers et al., 2010;Gazzola and Keysers, 2009).In the parietal and premotor nodes of this action observation network (AON), many of the voxels recruited by witnessing the actions of others are also activated Hypotheses about a particular (social) state, and its expected value, are built at a higher hierarchical level and transformed in more detailed sensory predictions that are sent to the lower hierarchical cortical level (feedback; purple arrows) via an internal model.At the same time sensory input enters the lower hierarchical cortical structures and is sent forward to higher levels (blue arrows).The content of each prediction and sensory input will depend on the hierarchical level the information is sent to (Bastos et al., 2012;Fukai et al., 2021).For instance hypotheses made in prefrontal regions will be more cognitive and abstract than those in primary sensory regions: if my belief is that the basketball player will soon bounce the ball on the floor, the prediction that the primary auditory cortex will receive will reflect low level auditory features of a ball hitting the floor.Similarly, inputs for primary visual cortices may include lines, while features in higher level visual cortices may represent entire houses, and my own lovely holiday home in more frontal regions.At each hierarchical level, the received prediction is compared against incoming sensory input, as illustrated in the figure by the summing junction symbol.This comparison creates the opportunity to generate a (prediction) error within the reinforcement (λ*PE) and error-based (λ*(αMSE/αθ) where MSE=mean square error) learning framework or an updating factor in Bayesian terms (p(θ)=p(input|θ)/p(input)), which are used to update the initial hypothesis at the higher level (green arrows), select the best outcome for a particular situation and over time, to learn.
Bayesian inference, Reinforcement and Error-based Learning describe these processes at different levels and are complementary rather than mutually exclusive (Spampinato and Celnik, 2021).The Bayesian framework describes the computations and mathematical formulation necessary to entertain, and update, the relative likelihood of all hypotheses (and inputs).The Reinforcement and Error-based learning framework refers to learning processes based on specific value or error estimations.At any given time, the Bayesian framework assumes a distribution of hypotheses, each with a given probability, while Reinforcement and Error-based learning, on the other hand, assumes that a single hypothesis about a value or variable of interest is maintained.
Hebbian learning describes the synaptic mechanism, based on optimal temporal contingency, by which synaptic feedforward and feedback inputs are selected and strengthened for each individual neuron (Bi and Poo, 2001;Hebb, 1949;Markram et al., 1997).Over time, it will reinforce connections that reflect the most frequent prediction-outcome associations, facilitating and supporting the predictive machinery and Bayesian computation at the synaptic level.
The nature of the input used to learn, what needs to be learned and the nature of the updating factor can differentiate the type of learning that occurs and the circuit that will be involved.For instance, classical reinforcement learning generally uses binary information (hit or miss) to learn which action or decision is successful and a prediction error as the updating factor, while error-based learning uses more continuous sensory inputs (e.g.how much strength was needed to throw the ball) to for instance improve the precision of movements and uses errors as updating factor.These differences determine which brain circuit will be primarily involved: for instance, classical value-based reinforcement learning mainly involves the striatal circuit, while error-based learning the thalamus-cerebellar pathway (Chase et al., 2015;Garrison et al., 2013;Taylor and Ivry, 2014).Beside value-based and error-based learning, other types of learning can occur and several mathematical formulations are proposed to describe different ways to compute the updating factor (Shakya et al., 2023).Again, the situation and what aspect of a skill needs to be learned will determine the primary computation performed in a particular brain circuit (Spampinato and Celnik, 2021).
These predictive and learning processes, which have been found to be powerful ways to select, predict and learn from our own behavior, can be adapted to predict and learn from the behavior of others.One of the main differences between applying these mechanisms to self and others is the fact that the sensory input might at times be missing in the case of others: e.g.we do not have direct access to other people's proprioception, nor to their emotional state.This might potentially impair the hierarchical chain of comparisons between predictions and inputs.The discovery of the mirror neurons system for actions (shaded light blue regions on the brain), and vicarious responses more in general, are thought to help while performing similar actions (Gazzola and Keysers, 2009), and in the monkey, these nodes contain single neurons that respond during both the observation and execution of specific hand actions, the so called mirror neurons (Fogassi et al., 2005;Gallese et al., 1996;Keysers et al., 2003;Kohler et al., 2002).While the AON is therefore well mapped, how these visual, parietal and premotor nodes exchange information to enable participants to perceive the intentions and goals of others from the movements we can observe has received comparatively little experimental attention.When we see someone's hand move towards a knife, how does our brain perceive that they want to grasp the knife to butter their sandwich?Traditional models have focused on a feedforward stream of information (Giese and Poggio, 2003) suggesting that neurons combining several receptive fields in visual cortices converge to create reaching sensitive neurons in STS that then trigger neurons in PFG and then F5, where activity of neurons also performing reaching and grasping in the observer become activated and thereby trigger a sense of intensionality normally associated with our own actions (Rizzolatti and Sinigaglia, 2016).
In contrast, Keysers and Perrett (2004) emphasized that reciprocal connections feeding information backwards from F5 to PFG to STS do not only exist, but could convey inhibitory predictions from premotor to visual cortices.To demystify how such predictions could be wired into the system, they proposed that while we view our own hand actions, the motor command in premotor cortices triggering the action precedes by some hundreds of milliseconds the contingent neural activity to the resulting sight of these actions in the visual cortices, so that Hebbian learning predicts the emergence of the predictive connections from premotor to visual cortices that have been empirically confirmed (Keysers et al., 2014;Keysers and Gazzola, 2014) (Box 1).Via these Hebbianly trained inhibitory feedback loops, the AON stops being a traditional feedforward recognition system and becomes a self-organizing dynamical system that infers motor goals from visual information, and then suppresses expected visual input from these inferred motor goals.Kilner and colleagues (Kilner et al., 2007) highlight how this architecture fits predictive coding with information from STS->F5 actually represents prediction errors, while the information from F5->STS represents predictions about sensory input given a certain motor goal or intention encoded in F5 neurons.Recognition of the intentions of others is then achieved when prediction errors from STS->F5 are minimal.They emphasize that premotor neurons also receive contextual information, such as where the actions are performed, that can help disambiguate the intentions behind an action.
Together these two lines of thinking provide a seductive hypothesis about how the brain learns to wire its connections across the visuoparieto-premotor action observation system to infer the intentions of others.Importantly, this view makes a number of testable predictions: the activity of premotor neurons encoding an action should be enhanced if contextual information predicts that action; should send sensory prediction backwards to parietal and visual neurons encoding that action; and activity in visual brain regions should be suppressed if premotor neurons can predict upcoming visual input.At the system level, these accounts predict that when sequences of actions can be correctly predicted, information should, somewhat counterintuitively, predominantly flow from premotor to visual brain regions -from inside out as it were.The traditional feedforward flow of information should in contrast be dominant when actions are surprising, be it because they are seen without context enabling predictions or, when they violate predictions.
That the action observation network overlaps so conspicuously with the system involved in controlling hand actions was a further source of inspiration to suspect its predictive architecture (Gazzola and Keysers, 2009;Kilner et al., 2007): Motor theorists have long held that an effective motor control systems must contain forward internal models predicting the sensory consequences of an action, and inverse models determining motor programs necessary to achieve a specific sensory consequence (Wolpert et al., 1998;Wolpert and Ghahramani, 2000;Wolpert and Kawato, 1998).Action observation could leverage the internal models of the observer to predict the actions of others, with a level of precision that increases with the similarity in their bodies (Gazzola and Keysers, 2009).
Below we will review the emerging evidence that supports these predictions and encourages us to view the action observation network as a predictive coding system.

Evidence from human fMRI and ECoG studies
While fMRI has been successful at localizing the nodes of the AON, testing the core predictions of predictive coding with fMRI has been hampered by the inability of traditional BOLD fMRI to reliably determine the direction of information flow (Smith et al., 2011): in electrophysiology, the direction of information flow is normally inferred from the relative timing of the neural activity of pairs of regions A and B, in contrast, fMRI measures the hemodynamic response that lags behind neural activity by delays that vary between brain regions and individuals (Handwerker et al., 2004), so that if A sends information to B, but B has a faster hemodynamic response than A, directional analyses may falsely conclude that region B sends the information to A. This is particularly problematic if the temporal delay in neural activity across regions is shorter than 100 ms (Schippers et al., 2011) -a significant problem given that the transfer between consecutive stages in a cortical hierarchy only takes about 20 ms (Schmolesky et al., 1998), and areas would thus need to be over 5 stages away from each other for their direction of information-flow to be reliably inferred using traditional fMRI.
A number of fMRI studies have nevertheless attempted to analyse the direction of information flow during action observation using Granger causality or dynamic causal modeling, and found data compatible with predictive coding.Schippers and Keysers found that when participants guessed the meaning of a sequence of gestures, the predominant direction of information flow was indeed in the premotor to middle temporal gyrus direction (Schippers and Keysers, 2011).Participants that show stronger premotor to parietal information flow are also better at recognizing biological motion (Sokolov et al., 2018).Several studies using dynamic causal modeling confirm that models that posit mutual connections between premotor and parietal, and parietal and visual nodes explain fMRI data during action observation better than those only positing bottom-up connections (Gardner et al., 2015;Sasaki et al., 2012;Urgen and Saygin, 2020).Finally one study showed that observing movements that are more familiar is associated with reduced information flow from visual to parietal regions (Gardner et al., 2015).However, given the above mentioned concerns about the reliability of inferences regarding directed information flow from BOLD signal, methods unaffected by the variability in the hemodynamic response function are needed.
Towards that effort, we filmed actors while they performed everyday sequences of actions, such as preparing a sandwich.Across several experiments, we then presented participants the individual motor acts (e.g.grasping a knife, cutting bread, etc) either in their intact, natural order to enable predictions, or in scrambled, randomized order, where they could not be predicted (Fig. 1A).In a first experiment using intersubject correlation, we showed that premotor and parietal nodes of the AON nodes have information about the sequence of actions that could be used overcome this shortcoming, by using our own experiences as a proxy for those of others (Keysers and Gazzola, 2014;Keysers and Perrett, 2004;Kilner et al., 2007).Again, the circuit will depend on the type of missing (e.g.somatosensory cortices will be recruited to simulate how much strength the other person is using) and available inputs, and the type of learning (value-based, error-based, etc)  for predictive coding (Thomas et al., 2018) (Fig. 1B).In a second experiment, using EEG, we found that activity in posterior electrodes likely to reflect visual activity, was reduced when viewing predictable acts (Fig. 1C), (Thomas et al., 2018), in line with the notion that predictions suppress early visual processing.In a third experiment, to measure the hypothesized increase in feedback activity from premotor to parietal nodes for predictable sequences without the limitations imposed by hemodynamic variability, we used 7 T fMRI to measure activity levels at several depths in the parietal AON node (Cerliani et al., 2022).This approach leverages that feedback connections from the premotor cortex reach the deep layers of the parietal cortex, while the feedforward visual information mainly reaches the middle layers (Gerbella et al., 2011;Shipp, 2007), allowing us to leverage the high spatial resolution of 7 T fMRI to measure feedback information flow rather than depending on the problematic temporal relationship between the hemodynamic response across regions as traditionally done at 3 T.We found that indeed, deep layers of the parietal lobe showed increased information about the actions of others when they are in their intact, predictable order, and that this information is likely to originate from premotor cortices, as shown using intersubject functional connectivity (Fig. 1D) (Cerliani et al., 2022).
In a final experiment, to more directly explore neural activity, we showed these movies to epileptic patients with EletroCorticoGraphic (ECoG) electrodes covering pairs of nodes of the AON.Unlike EEG that measures electrical activity from the scalp, ECoG measure activity locally, in the ~3 mm around each electrode (Dubey and Ray, 2019) and with higher signal-to-noise ratio across a wide frequency range (Kanth and Ray, 2020), including both β frequencies associated with feedback information from premotor, and high-γ frequencies associated with feedforward information from visual nodes (Fries, 2015;van Kerkoerle et al., 2014).We found that when acts can be predicted (intact sequences) compared to when they cannot (scrambled sequences), feedback information flow from premotor to parietal nodes is indeed increased in the high-β range (20-30 Hz) and feedforward information flow from visual to parietal nodes is reduced in the high-γ range .Interestingly, the premotor feedback increases before the beginning of the next act (Fig. 1H,I), when they are predictable, whilst the visual feedforward information increases when acts are unpredicted just after they appear (Fig. 1F,G).Such a pattern is reminiscent of the relationship between FEF and visual cortices during a saccade, with top down activation boosting the activity of visual neurons in peripheral locations that will fall into the fovea after the saccade (Rao et al., 2016).

Evidence from human TMS and tDCS studies
That the motor system is involved and necessary for predicting the actions of others is also borne out by transcranial magnetic stimulation (TMS) studies.Applying a TMS pulse over the hand representation of M1 triggers a measurable twitch in hand muscles -the so-called motor evoked potential (MEP).Imagining or observing an action involving a particular hand muscle is known to increase the magnitude of this MEP measured from the corresponding muscle, thereby providing an objective measure of otherwise subthreshold motor activation in the brain at the moment at which the pulse is delivered (Naish et al., 2014).Interestingly, while viewing snapshots suggesting a finger action such as flicking a ball, the maximal MEP facilitation occurred for snapshots taken ~200 ms before the phase of maximal muscle contraction of the actor, suggesting that the motor system was anticipating the actors actions by ~200 ms (Urgesi et al., 2010).Such anticipation may be particularly important for elite athletes.Indeed, elite basketball players can predict whether another player's free shots will land in the basket ~200 ms earlier, and more accurately, than basketball coaches that have extensive visual expertise but less motor expertise with the observed actions (Aglioti et al., 2008).These superior predictions could be made by players before the ball had even left the hand of the observed player, when the angle of the little finger was the only visual cue to the direction of the ball.MEPs from the abductor digiti minimi muscle (ADM) that controls the corresponding little finger in the observer could significantly predict whether the observed shot would, 1 s later, enter the basket or not, demonstrating that the motor system indeed has the signals to anticipate the future of other people's actions -as long as one has, like elite players do, a motor program closely matching that of the people we observe.That the motor system contains signals that can predict the actions of others does not demonstrate that people actually rely on such motor signals to perceive the future actions of others.To address this causal question, TMS can also be used to disrupt premotor activity using repetitive TMS.Perturbation of the dorsal premotor cortex has been found to impair the ability of football players to predict whether a penalty would fly right or left of the goal keeper (Makris and Urgesi, 2015) as well as the ability of participants to determine whether the continuation of an action that has been temporarily occluded was offset in time (Brich et al., 2018;Stadler et al., 2012).Also altering activity in the ventral premotor cortex using transcranial direct current stimulation alters the ability to anticipate which of two objects a reaching movement would ultimately gasp (Avenanti et al., 2017).Together, this confirms the notion that the premotor regions play a causal role in predicting the actions of others (Keysers et al., 2018).

Evidence from monkey work
Several single cell recording studies in macaques further support the notion of predictive coding during action observation using cellular resolution techniques.Umilta et al. (2001) recorded from premotor mirror neurons that responded when the monkey grasped an object and when viewing others grasp objects.These neurons responded significantly less when viewing a demonstrator mime a similar grasping without an object, confirming that their activity contains information about the goal of the action -grasping that object.They then tested these Fig. 1.Evidence for Predictive coding in experiments using sequences of actions.(A) Example frames from a movie of making a butter and jam sandwich either in the intact (green throughout the figure) or scrambled (black) order, with the intact but not the scrambled sequence enabling predictions.Note: to ensure that both intact and scrambled movies have similar levels of visual discontinuity, actions were recorded using two cameras 45 deg apart, and adjacent acts were always from different perspectives.In the intact condition, the action however continued across the perspective change, as if often the case in professional movie making.(B) Brain regions showing higher intersubject correlation (ISC) for intact sequence, thereby showing evidence of containing information about the probable sequence of acts in natural sequences.Note the presence of the parietal and premotor node of the AON.(C) EEG activity from posterior electrodes (shown in black on the head outline) aligned to the camera change between every two acts (vertical line) for intact and scrambled sequences, showing the suppression of predictable visual input over occipital electrodes.(D) Depth resolved ISC values, with positive values indicating higher information content for intact sequences, and stars indicating the significant increase in information in deep layers.The grayscale photograph is a histological image of the supramarginal gyrus with its 6 layers, and the blue gradients illustrate the layers receiving premotor feedback and visual feedforward information from high-level visual areas.(E) Illustration of the ECoG electrodes from 9 patients, with each dot color indicating electrodes from the same participant.AON regions of interest are circled.Letters refer to the panels showing the power or direction of information transfer.The light blue / lilac colored arrows match the directions of information flow in G and H (F) High-γ power relative to the onset of each act in the visual electrodes is significantly suppressed (red bar) for intact sequences (green) compared to scrambled sequences (black).(G) The dominant direction of high-γ activity was from visual to parietal for scrambled sequences (black), but not intact sequences (green).(H) High-β power was increased just before and shortly after the onset of each act in the intact sequences, but not the scrambled sequences, and was (I) in the premotor to parietal direction for the intact but not the scrambled sequences.(F-I) red bars mark periods of significant difference.Results for Intact sequences are always shown in green, scrambled sequences in black.
neurons while the monkey saw an experimenter's hand disappear behind an occluding screen behind which the monkey had seen the experimenter place the same object or not.At the moment when the hand disappeared behind the screen, the population of mirror neurons responded more in trials in which the monkey could infer that the hand was grasping an object, despite then identical visual input: a hand disappearing behind an occluder.Premotor mirror neurons therefore supplement bottom-up visual information with prior knowledge to generate what could best be described as the hypotheses about observed actions that predictive coding posits the premotor cortex to generate.Kohler et al. (2002) tested premotor mirror neurons that responded to the sound of particular actions, e.g.ripping paper or breaking a peanut, and found if the monkey could see a human prepare to perform these actions, responses anticipated the actual ripping or breaking by several hundreds of milliseconds.The activity in premotor neurons therefore predicts the occurrence of a particular action based on preceding observed movements.Fogassi et al. (2005) trained monkeys to grasp food to either eat it, or place it into a container, and recorded from parietal mirror neurons that responded differently during grasping based on the goal (eat or place).When observing an experimenter repeatedly grasp to place versus to eat, these neurons also differentially responded to the observation of grasping, thereby anticipating the differential goal.Together these studies confirm that both premotor and parietal mirror neurons predict upcoming actions based on the recent past, in line with the above-reviewed human data that both of these regions have information about the sequence of actions.Later Maranesi et al. (2014) developed a paradigm testing whether premotor neurons can use arbitrary cues to predict upcoming motor actions.Monkeys were trained that at the end of a high-pitched tone, they are rewarded for grasping an object, and at the end of a low-pitched tone, they are rewarded for refraining from grasping.They showed that the populations of mirror neurons in F5 showed differential predictive activity when witnessing an experimenter perform a similar task, significantly higher following the high-than low-tone 340 ms before the experimenter actually started to move in the high-tone trials.Using a similar paradigm, Ferroni et al., (2021) compared the activity of mirror neurons in the parietal (AIP) and premotor cortex (F5), and neurons to predict whether a demonstrator would grasp much earlier in F5 (260 ms before grasping) than AIP (40 ms before grasping).These two experiments show that the AON can use arbitrary cues to predict actions by several hundreds of milliseconds, and supports the notion that such predictions originate within the earlier responding premotor nodes and are then transmitted to the later responding parietal AON neurons.

The AON as a predictive coding network
For a long time, action observation had been studied by measuring brain responses to brief, 1-2 s motor acts, such as grasping an object, presented in isolation.Seeing or hearing such individual acts can suffice to trigger robust responses that are tightly locked to the timing of that act in monkeys (Gallese et al., 1996;Keysers et al., 2003;Umiltaà et al., 2001) and humans (e.g.Caspers et al., 2010 for a review; or Gazzola and Keysers, 2009 for an experimental paper).However, in real life, we seldom observe motor acts so out of context: we often see them embedded into longer meaningful sequences aimed at achieving intentions such as preparing a meal.Increasingly, data shows how, although neurons in the AON may respond to individual acts, as a dynamic network, the AON appears to be a predictive coding system with premotor and parietal regions generating hypotheses about future acts that are sent via feedback connections to visual nodes where they suppress expected visual consequences.Rather than passively reacting to the social world from the outside-in, our brain actively predicts it, with perception becoming an active process from the inside-out.

Predictive coding of the emotional state of others
For social species, being able to predict the emotional states of conspecifics would seem to have high value.Predicting what will upset or please our teacher is an important learning signal that could improve our performance.When facing decisions that have consequences for others, correctly anticipating their hedonic response would be helpfulbe it for choosing what Christmas present will please our partner, or what barb will hurt our opponent.Sharing a conspecific's distress in a particular context could also provide valuable information to predict danger and engage in appropriate defensive action (Keysers et al., 2022).Finally, given that social interactions involve a mutual influence, with the emotional states from all parties continuously influence each others (Bachmann et al., 2022;Han et al., 2019;Thornton and Tamir, 2017), anticipating the affective state of others and how they will influence our own would be critical to maintaining our own emotional state within desired bounds.
So do mammals process and predict the emotions of others and how?Many human neuroimaging studies have revealed that as for action observation, neuronal substrates (in particular the anterior insular and anterior cingulate cortices) critical for our own emotions become reactivated by the same emotional state witnessed in other individuals (Jauniaux et al., 2019;Lamm et al., 2011;Morelli et al., 2015;Singer et al., 2004;Wicker et al., 2003).Recording from neurons in the anterior cingulate of rats and mice revealed that they contain a mix of neurons only responding when the animal itself is in pain, neurons only responding when witnessing another in pain, and some, called pain mirror neurons, respond to both pain in the self and other (Carrillo et al., 2019;Zhang et al., 2024), with preliminary evidence for a similar mix existing in electrophysiological recording in humans as well (Hutchison et al., 1999).As theories of emotions increasingly consider emotions as intrinsically predictive (Barrett, 2017;Critchley and Garfinkel, 2017;Seth, 2013;Seth and Friston, 2016), that the very regions suspected of predicting future affective states of the self are recruited while witnessing the emotions of others conspires to beg the question of whether they also predict future states of others (Gendron and Barrett, 2018;Ishida et al., 2015) (Box 1).

Learning paradigms are critical for investigating predictive coding for the emotions of others
As mentioned above, gathering experimental evidence for hierarchical predictive coding architectures from human fMRI is problematic if attempting to leverage fine temporal delays between brain regions.EEG and TMS are also limited in their utility for investigating affective processes, as the core regions, such as the insula and cingulate are too deep for these techniques.Accordingly, unlike for the case of action observation, where much of the evidence for predictive coding stems from studying hierarchical predictive coding and the direction of information flow, for emotion observation, studies have explored how brain signals change from trial to trial as animals and humans can learn to predict outcomes of emotional value, specifically rewards and punishments, for others.The rationale is that theories of learning make two quantitative predictions about how brain signals should change over trials if the brain engages in decisions based on predicted hedonic values (i.e.rewards -punishments) and learns from prediction errors (i.e. the difference between predicted and actual outcomes) (Dayan and Daw, 2008;Rescorla and Wagner, 1972;Schultz, 2002) (Box 1).First, predictions of outcomes should be already visible during a decision-phase before outcomes are revealed.Second, if the brain simply processes outcomes, witnessing a conspecific receive an unpredicted or a predicted reward should produce identical responses in the brain, if it predicts, these responses should be rather different, and proportional to the predictions or prediction errors made by learning theory.

Dopaminergic signals for rewards to others
For reward to the self, it has been widely shown that midbrain dopaminergic neuronal activity codes reward predictions and prediction errors (e.g.Bayer et al., 2007;Bayer and Glimcher, 2005;Diederen and Fletcher, 2021;Morris et al., 2010;Schultz, 2016;Waelti et al., 2001).This has placed dopamine at the center of predictive coding for rewards.Whether dopaminergic activity also tracks reward expectations for other individuals is less clear.Kashtelyan et al. (2014) trained rats that the presentation of a light would predict reward for the self.Subsequently they moved the trained rat to an adjacent compartment, from which it could observe a conspecific receive the expected reward after the appearance of the same visual cue.Voltametry from the ventral striatum during the first trial of the test session shows dopamine release during the presentation of the cue and at reward delivery.Critically, the response was larger than if there was no conspecific in the reward compartment, indicating that the dopamine signal predicted the reward delivery to a conspecific rather than the delivery of a reward in the other compartment per se.That dopamine was released already at the appearance of the cue supports the idea that striatal dopamine release may indeed predict positive, emotionally relevant outcomes for others.What is less clear is why such anticipatory response diminished so quickly over trials.Although direct measures of dopamine in humans can be done (Bang et al., 2020), they remain rare, and evidence for the involvement of dopaminergic activity in vicarious reward in humans, remains indirect and mainly comes from the observation that fMRI striatal activity is elicited both by rewarding events to self and other individuals (e.g.Mobbs et al., 2009;Monfardini et al., 2013;or Morelli et al., 2015 for a meta-analysis).Striatal responses to the reward of others also depends on the closeness of the relationship with the other (Brandner et al., 2021;Mobbs et al., 2009).Whether the human striatal signal is predictive remains less clear.

Anterior cingulate cortex predicts rewards to others
Monkey work on vicarious reward offers further insights.In this context, one monkey learned the consequences associated with three different cues: one would predict juice for the monkey themself, one to another monkey, one for no-one.Then two of these symbols were presented at the same time for the monkey to choose from.ACC single neuron recordings show that when the monkey chose to deliver the reward to the other conspecific, some neurons started to respond to the predicted reward for the other before the reward was delivered (Chang et al., 2013).Lesion of the ACC stopped the monkey from learning which symbol predicts reward to the other (Basile et al., 2020), suggesting the necessity of the ACC in vicarious reward learning and prediction.That the ACC may be involved in the prediction of reward to others is further supported by the work of Lockwood and colleagues (Lockwood et al., 2015) in humans.In that study activity in the ACC, particularly in a subregion corresponding to areas 24a′/24b in rodents (Vogt et al., 1995), correlated with the probability of another individual to receive a reward already at the time of the anticipatory cue presentation.

Anterior cingulate cortex and amygdala predict footshocks to others
For punishment or pain, rodents that have observed a conspecific receive footshocks in a particular environment will freeze, both while witnessing the demonstrator receive shocks and, the next day, when placed back in that environment (Atsak et al., 2011;Carrillo et al., 2019;Jeon et al., 2010;Keysers et al., 2022) showing that they learn to predict emotionally negative events in the future.This effect depends on the integrity of the ACC (Atsak et al., 2011;Carrillo et al., 2019;Jeon et al., 2010;Keysers et al., 2022), which synchronizes its activity with the basolateral amygdala (BLA) in the theta-frequency range during learning.In a similar paradigm, witnessing another mouse receive footshocks at the end of a particular sound leads the witness to later freeze when hearing that sound (Allsop et al., 2018), and neurons in the witness' ACC and BLA, which initially only responded to witnessing the demonstrator receive shocks, started to respond to the cue as learning progresses.This suggests that these neurons encode a prediction of future punishment, and the connections from the ACC to the BLA were necessary for such predictions to be acquired.Complementing these results Silverstein and colleagues (Silverstein et al., 2024) showed that mice dorsomedial prefrontal (dmPFC) neurons (possibly partially overlapping with the ACC region from Allsop et al.'s work) are necessary for learning to freeze when hearing a tone paired with footshocks to a conspecific and that a subset of dmPFC neurons responding to shock observation also started to show predictive responses during tone-presentation as learning progresses.Interestingly, dmPFC neuron projecting to the ventrolateral periaqueductual gray (vlPAG), and not those projecting to BLA, were involved in the acquisition of vicarious fear.
Together, these animal studies suggest that a network involving several medial frontal regions together with the amygdala and vlPAG can learn to anticipate the punishments of others.As these regions are also necessary for the freezing behavior of observers in response to predictive cues, they may serve to use the experiences of others to predict threats to the self (Keysers et al., 2022).Whether these signals are thus strictly predicting the emotions of others, or predicting threats to the self remains unclear.
In line with Allsop's results, human fMRI work on how participants learn about threats from witnessing the painful experiences of others also shows the amygdala to be activated during the presentation of cues predicting a shock to another person, both during the learning (presentation of the cue followed by the other getting shock) and test phase (presentation of the cue alone) (Olsson et al., 2007).Building on this seminal work, Haaker and colleagues additionally report the opioid system mediating threat responses in the amygdala and PAG during the social fear acquisition and longer lasting fear responses (Haaker et al., 2017).

Cingulate predictions may guide moral decision-making
We are often faced with the need to decide whether we prefer an action that benefits ourselves more but harms others and one that benefits ourselves less but prevents harm to others.Such moral decisions require us to predict the consequences of our actions on others, and the type of anticipatory signals found in the ACC would be well poised to support such decision-making.Indeed, rats and mice can learn to avoid using a lever to obtain food if doing so delivers a footshock to others (Greene, 1969;Hernandez-Lallement et al., 2020;Song et al., 2023).Interestingly, the ACC that responds to the pain of others both in rodents and human research (Carrillo et al., 2019;Jauniaux et al., 2019;Lamm et al., 2011;Zhang et al., 2024) and has shock predicting signals (Allsop et al., 2018;Silverstein et al., 2024), is necessary for rodents to decide against actions they anticipate to harm others (Hernandez-Lallement et al., 2020;Song et al., 2023).Several human fMRI studies reveal signals in the ACC that could serve a similar decision-guiding anticipatory role.ACC activity is found to correlate with interindividual differences in the value-update for vicarious pain (Lengersdorff et al., 2020), with the subgenual ACC to dorsolateral prefrontal cortex functional connectivity correlating with the decision to change actions to prevent harm (Lockwood et al., 2020).These results were further supported by the work of Fornari and colleagues (Fornari et al., 2023), in which the ACC was similarly found to correlate with updating the predicted pain to others in a morally conflicting situation.Finally, in Lockwood at al., (2020) the ventral striatum was found to correlate with the pain prediction error for both self and others, and a cluster encompassing the thalamus and caudate nucleus to correlate with the vicarious pain prediction error, suggesting complementary brain regions contribute to predict vicarious pain and its potential consequence for the self and others.A later study from the same author further investigated the role of the ACC in vicarious reward anticipation by asking participants to learn which of two symbols would most likely result in a monetary reward for self or someone else (Lockwood et al., 2016).They found that a more ventral subregion of the ACC, the subgenual anterior section in particular, correlated more strongly with the vicarious reward prediction error signal than with the reward for self.Finally, Apps and colleagues show that activity in the ACC of teachers also reflects the prediction of students to err in a task (Apps et al., 2015).Together these studies confirm that also in humans, regions along the frontal midline play a role in guiding behavior based on the predicted emotional consequences of our actions, be they expectations of reward or punishment.

Conclusions
Traditionally the field had studies how the brain processes the facial expressions of the emotions of others by presenting faces in randomized order at randomized intervals out of context to isolate the neural mechanisms responsible for processing the emotions of others without the interference of predictions.The traditional approach additionally expected emotional states to activate a relatively fixed brain circuit across participants and context.If we accept the idea that different input from the environment could differentially contribute to different predictions tuned to the current situation, the field of (vicarious) emotions might call for an analytical approach more appropriate for identifying more dynamic predictive systems (Lee et al., 2021).However, over the past decade, embracing learning paradigms in which human and animal participants can learn to predict that certain stimuli or actions can lead to rewards or punishers to others have started to reveal that many of the structures associated with our own emotions actually carry signals that can anticipate the emotions of others and compute prediction errors that support learning when outcomes differ from those predicted (Joiner et al., 2017).Whether these signals serve primarily to predict the state of others, or the value this has for the self -either in terms of predicting self-directed threats, or to guide moral decision-making, remains poorly understood.

Theoretical approaches to the predictive processing account of autism
The mounting evidence for the presence of predictive coding for the actions and emotions of others in humans, monkeys and rodents begs the question of whether embracing the importance of predictive coding could help clinicians better understand the etiology of atypical social functioning in their patients.Here, as an example, we will review how predictive coding is playing an increasing role in the study of Autism spectrum disorder (ASD).ASD is a neurodevelopmental condition marked by core impairments in social interaction, communication, and restricted or repetitive behaviours and interests (Diagnostic and statistical manual of mental disorders: DSM-5TM, 5th ed., 2013).People with autism are usually described as experiencing difficulties in social learning, understanding social cues and, or engaging in reciprocal interactions.Additionally, sensory atypicalities (i.e.differences in how individuals with autism process and respond to sensory stimuli from their environment) are often reported.These atypicalities can involve hyper-or hyposensitivity to sensory input, of one or more of our senses: sight, sound, touch, taste, smell, and proprioception.
In recent years, a predictive processing account of autism has been put forward as a possible unifying framework able to explain its diverse clinical manifestation (Palmer et al., 2017;Pellicano and Burr, 2012;Van de Cruys et al., 2014).Since its first appearance, the predictive processing account of autism has been revised to accommodate more and more complex explanations of the condition and its behavioral expression.The simple bayesian model proposed by Pellicano and Burr (2012) proposes that individuals with ASD exhibit a higher learning rate in perceptual inference, which influences how they weigh prediction errors against sensory signals.This higher learning rate is associated with heightened sensitivity to sensory stimulation, as evidenced by hypersensitivities to sensory input and sensory avoidance behaviors observed across various perceptual modalities.Multiple evidence of reduced susceptibility to certain sensory illusions and decreased effects of perceptual adaptation (see Palmer et al., 2017 for a comprehensive review), seems to corroborate the idea of a higher learning rate and a reduced influence of prior sensory history on current perception.This simple model, though, was unable to fully explain the heterogeneity observed within the condition and not all studies have found support for the proposed mechanism (Palmer et al., 2017).
Therefore, a broader precision modulation model attempted to better capture the characteristics of autism (Lawson et al., 2014).This framework proposes differences in context-sensitive adjustment of precisions rather than persistently higher weighting of prediction errors (Arthur et al., 2023).Sensory information integrates with expectations developed over both longer and shorter timescales in a hierarchical fashion.Consequently, context or volatility becomes a central element to consider when evaluating probabilities of events.For example, it has been suggested that adults with ASD distinguish less between expected and unexpected outcomes, pointing towards reduced surprise in a probabilistic associative learning task, both at the behavioural and physiological level (Lawson et al., 2017).This has been explained through the tendency of participants with ASD to overlearn about the volatility when an uncertain environment is present, which might lead to less learning concerning probabilistically deviating events, giving rise to the perception of unexpected events as less surprising (Lawson et al., 2017).Therefore, individuals with ASD tend to struggle in situations with high contextual ambiguity, such as social situations (Palmer et al., 2015).Notably, learning deficits due to a failure to integrate social context to adapt one's belief precision can be observed in the general population, depending on high autistic traits (Sevgi et al., 2020).However not all studies have found support for the aberrant precision hypothesis advocated by Lawson and colleagues (Manning et al., 2017;Ward et al., 2022).One possible reason for the mixed findings is that imbalances between priors and likelihoods could be produced by a variety of underlying mechanisms, which current studies, due to difference in definition methodologies and sample sizes, were not able to fully understand.Guidelines for future research have been recently proposed (see Angeletos Chrysaitis and Seriès, 2023 for a critical discussion).

Autism through the lens of active inference
We can also try to understand autism from the standpoint of active inference, i.e. reducing possible prediction errors by acting on the world to bring sensory input closer to predicted outcomes.The balance between action and perception, mediated by the context-sensitive estimation of precisions, is crucial for how individuals interact with and sample the world around them, and accordingly, such balance plays a central role in autism as well.This involves, for example, optimizing internal models by determining where and for how long individuals sample the visual field, as well as how they interact with others to understand their mental states.One key aspect of ASD is prolonged sensory sampling behavior due to reduced sensory attenuation, which leads to an excessive focus on certain stimuli or objects (Lawson et al., 2014).Children with ASD often exhibit unusual fixation on specific stimuli, such as bright lights or repetitive patterns.Additionally, they may engage in repetitive visual exploration behaviors, examining fewer objects but in more detail.These tendencies are considered early signs of ASD and are included in diagnostic criteria as "unusual interest in sensory aspects of the environment.".Motor initiation is also affected in individuals with ASD, as evidenced by slower reaction times and atypical movement preparation (Palmer et al., 2017).Repetitive or stereotyped behaviors, such as rocking or hand flapping, are common in ASD and can be observed as early as infancy.These behaviors serve as a way to continually sample sensory information, reducing uncertainty in the brain's representation of the environment.Similarly, socially pertinent behaviors, such as reduced orienting to social stimuli and impaired joint attention, can also be understood in terms of sensory sampling for perceptual inference (Palmer et al., 2017).Sensory-sampling behaviors, motor initiation, repetitive movements, and social interaction in ASD are interconnected and can be understood in terms of the balance between action and perceptual updating.For example, difficulties in sensory processing (sensory-sampling behaviors) can affect motor initiation and lead to increased repetitive movements as a coping mechanism.These challenges can, in turn, impact social interactions, as the individual may find it hard to engage with others while managing sensory overload and motor difficulties.Overall, understanding these behaviors in terms of the balance between action and perceptual updating provides a comprehensive framework.It highlights that the difficulties experienced by individuals with ASD are not isolated issues but are part of a dynamic system where sensory processing, motor activities, repetitive behaviors, and social interactions constantly influence each other.Improving the balance between these elements can potentially lead to better interventions and support strategies for individuals with ASD

Autism and the predictive processing account of emotions
In the emotional domain, recent theoretical work indicated predictive processing alteration in interoception as a possible mechanism of the manifestations and development of the socio-emotional atypicalities in autism (Djerassi et al., 2021;Seth and Friston, 2016).Interoceptive inference, in the context of predictive processing, refers to the brain's process of generating predictions about the body's internal state, also known as interoception (Barrett and Simmons, 2015;Craig, 2002).It involves making continuous predictions and updating these predictions based on incoming sensory information from within the body, including factors like heart rate, blood pressure, temperature, hunger, etc (i.e.allostasis).This iterative process of generating predictions, comparing them to incoming sensory data, and updating the internal model helps the brain maintain a more accurate and up-to-date representation of the body's internal state, perception of bodily sensations, and perception and regulation of emotional experiences.According to (Seth and Friston, 2016), an inability in autism to place interoceptive cues, triggered for example by interactions with the caregiver, into the proper context prevents the accurate recognition of the source of interoceptive effects arising from social interactions.To put it simply, if the precision of prediction errors related to interoception is not properly adjusted, it could result in autistic infants being overly sensitive to interoceptive signals (referred to as autonomic hypersensitivity).This hypersensitivity would have significant implications for the development of the typical associations between emotional states and external sensory cues.Consequently, individuals might have difficulty distinguishing their own emotional state from that of others, struggle to attribute significance to external cues like a mother's facial expressions, and have less well-defined self-representations, all manifestations often referred to autistic individuals (Quattrocki and Friston, 2014).A similar account is described by Djerassi et al. (2021).In their work, the authors argue that autism is characterized by a general variation in allostasis-driven learning, which in turn adversely affects the rewarding experience of social interactions.Mothers, or the primary caregivers, stand out as not only the most consistent source of information in the infant's environment but also the most attention-grabbing and potentially rewarding figure due to their crucial role in maintaining the infant's overall well-being (allostasis).The strong regularity between presence of the caregiver and allostatic processes enhances social motivation and strengthens the acquisition of social concepts and behaviors.As caregivers take charge of regulating the infant's allostasis, infants not only develop fundamental perceptual concepts but also more abstract notions, such as understanding emotions and mental states (Atzil et al., 2018).According to this model, autism results from an atypical rate of learning, due to variation in domain-general processes of perception, motor control, allostasis regulation, or their multi-modal integration into concepts.Variation in any of these processes can manifest as atypical social development, which can result in reduced social orienting, social seeking, social liking, and deficient social learning, central characteristics of the condition.Empirical evidence of predictive processing alterations in the emotional domain is so far limited.Keating et al. (2023) observed that autistic individuals, despite more precise visual emotion representations, did not show enhanced accuracy on a task which indexed emotion recognition from dynamic stimuli, suggesting that autistic individuals may not use their (precise) emotion representation to help them recognize emotional expressions.This finding aligns with Bayesian accounts of autism which posit that autistic individuals are less influenced by priors than non-autistic people counterparts.On the other hand, Finnemann et al. (2021) observed that self-versus externally-generated action effects in terms of sensory attenuation do not differ between adults with and without autism.Indeed, autism was associated with normal levels of sensory attenuation of internally-generated force, indicating intact low-level proprioceptive prediction abilities.In general, it is important to note that atypical predictive processing does not always result in a decline in performance, rather more subtle differences become evident only on implicit measures, such as brain activity (Chiappini et al., 2024;Sapey-Triomphe et al., 2023).In the following we will review evidence of prediction processing alteration at the neurophysiological level.

Autism, predictive processing and neurophysiological evidence
Altered functional brain connectivity is often reported in ASD and it has been suggested as the neural manifestation of predictive processing deficits.However, findings of hyper-and hypo-connectivity have been mixed (Cerliani et al., 2015;Holiga et al., 2019;Lau et al., 2019;Supekar et al., 2013).Whilst some studies suggest that hyper-connectivity is more common in children with ASD on whole brain as well as subsystem level (Supekar et al., 2013), others found local underconnectivity, specifically in the dorsal posterior cingulate cortex and in the right medial paracentral lobule (Lau et al., 2019).Moreover, Holiga et al., (2019) provided evidence for hypoconnectivity, primarily in sensory-motor regions, and hyperconnectivity in prefrontal and parietal cortices.Although the literature therefore converges on the presence of altered connectivity, the details of how it is altered remains poorly understood.Recent computational-fMRI studies have found reported differences in prediction-error representation in associative learning tasks in ASD (Sapey-Triomphe et al., 2023): within a changing environment, neural encoding of predictions were similar in neurotypical and ASD participants, differences between the groups emerged in the anterior cingulate cortex and putamen with regard to the processing of prediction errors.
Recently, we showed that when anticipating social and nonsocial rewards, autistic individuals report typical levels of wanting and effort to obtain the desired stimuli.Such anticipation though is underpinned by an overprocessing of the stimulus information as indicated by increased activity and connectivity between reward related areas and primary sensory and visual areas, compared to neurotypical individuals (Chiappini et al., 2024).Notably, such overprocessing, which does not result in behavioral differences in such a controlled and repetitive task, may show its effects on more complex and uncertain situations of reward selection and decision making, such as the one encountered in everyday life, leading to the well-characterized deficits observed in this condition.

Conclusion
Given the complex and sometimes contradictory evidence in the literature, it is fair to say that whether ASD should be understood as a deficit in predictive coding remains unclear.However, the predictive processing account of autism has proven to be an influential framework that has energized research into the diverse clinical manifestation of ASD and its individual heterogeneity.It generates hypotheses about how individuals with autism perceive and interact with their environment that can be tested experimentally.It holds promises for the identification of biomarkers that could aid in early diagnosis or prognosis.Novel therapeutic interventions tailored to the specific sensory processing differences observed in autism can be informed by predictive processing research.These therapies could focus, for example, on modulating prediction errors, recalibrating sensory precision, or promoting adaptive predictive mechanisms.
By considering the variability in predictive processing profiles among individuals with autism, clinicians could develop personalized treatment plans that address each individual's unique sensory processing challenges.In summary, the predictive processing account of autism offers a source of inspiration for understanding the cognitive, affective and neural mechanisms underlying the condition.

Overall conclusions and future directions
The studies summarized in this review demonstrate an emerging awareness of the importance of predictions in social cognition and a nascent body of evidence confirming that the brain indeed carries signals that contain predictive information about the upcoming actions and hedonic states of others.This body of evidence comes in part from a change in the experimental paradigms used to probe social cognition.Traditionally, the visual perception of the actions and emotions of others had been measured using brief stimuli in the order of a few seconds presented in randomized order.This approach was motivated by the idea that social perception is rapid, and can therefore be studied with brief stimuli, and that preventing predictions by randomizing the order of stimuli would provide the cleanest insights into how the brain processes social signals.As neuroscientists became increasingly seduced by notions of the brain as Bayesian predictive machines, some dared to study brain functions in paradigms that permit predictions.Movies of actions are for instance embedded into meaningful sequences and the emotions of others become predictable through classical or operant conditioning.Doing so has unveiled the existence of significant feedback information flow in the action observation network and neural activity that comes to predict emotionally meaningful outcomes to others.In addition, rather than simply processing the actions and emotions of others, once these stimuli become predictable, several brain regions have activity that correlates better with prediction error than raw outcomes, becoming increasingly attenuated as the outcomes become more predictable.In addition to embracing new paradigms that can reveal predictive coding, the field of social neuroscience had to leverage a multitude of methods to start providing evidence for predictive coding.Animal studies in which neural activity at cellular resolution can be measured at high temporal resolution are uniquely poised to provide evidence for activity representing the actions and emotions of others before they occur.Animal studies also provide the opportunity to interfere with deep brain regions with high temporal and spatial resolution to provide evidence for the importance of specific circuits in generating predictions and in influencing the sensory representations in earlier brain regions.In humans, ECoG and depth resolved fMRI can provide critical insights into the balance of feed-forward and feed-back in ways that complement the animal studies.It will take decades and the willingness of social neuroscientists to unleash these methods systematically in paradigms in which the emotions and actions of others are predictable, to bring the field from the hints of predictive coding we review here, to a solid understanding of the predictive coding architecture of the brain and its potential dysfunctions in disorders of social cognition.However, this effort promises a conceptually profound shift in our understanding of the social brain from a system that processes the states of others to one that generates an active inference of our social future.
used in that circumstance.caption on next page)