Integrating the global neuronal workspace into the framework of predictive processing: Towards a working hypothesis

Integrating the global neuronal workspace into the framework of predictive processing: Towards a working hypothesis

Consciousness and Cognition 73 (2019) 102763 Contents lists available at ScienceDirect Consciousness and Cognition journal homepage: www.elsevier.co...

452KB Sizes 0 Downloads 25 Views

Consciousness and Cognition 73 (2019) 102763

Contents lists available at ScienceDirect

Consciousness and Cognition journal homepage: www.elsevier.com/locate/concog

Review article

Integrating the global neuronal workspace into the framework of predictive processing: Towards a working hypothesis

T

Christopher J. Whyte School of Psychology, University of Sydney, NSW, Australia

ARTICLE INFO

ABSTRACT

Keywords: Predictive processing Predictive coding Limbic workspace Consciousness Conscious access Global neuronal workspace

Hohwy (2013) proposed an account of conscious access that integrates the global neuronal workspace (GNW) into the framework of predictive processing, a view that I term the predictive global neuronal workspace (PGNW). Whilst promising, the PGNW is theoretically underdeveloped and empirically underexplored. The aim of this article is to outline the empirical predictions that distinguish the PGNW from other workspace models. I do so by (i) placing the PGNW in close contact with experimental work and cashing out a set of predictions that distinguish it from the standard formulation of the GNW, (ii) exploring the evidence for the first of these predictions in the context of bistable perception and the conscious processing of auditory regularities, and (iii), contrasting the PGNW with Chanes and Barrett (2016) limbic workspace. Combined, these arguments show that the PGNW is both testable and supported by current evidence.

1. Introduction Predictive processing (PP), aims to subsume perception, action and cognition under one computational umbrella: a Bayes approximate process of prediction error minimisation. According to this view, what the brain ultimately does is minimise the difference between predictions and sensory signals. Using an internal model of the sensorium, the brain generates cascades of top-down predictions that meet bottom-up signals at each successive level of the cortical hierarchy (Friston, 2005; Clark, 2016; Hohwy, 2013). In contrast to standard feature detection models that view perception as a predominantly feed-forward phenomenon consisting of hierarchical feature extraction, PP proposes that it is only the mismatch between top-down predictions and bottom-up input (the prediction error) that is fed-forward through the hierarchy. In recent years, PP has been applied to psychopathology (e.g. Fletcher & Frith, 2009), attention (Hohwy, 2012; Kok, Rahnev, Jehee, Lau, & de Lange, 2011), action (for a review see Clark, 2013) and a diverse array of perceptual phenomena (Hohwy, Roepstorff, & Friston, 2008; Rao & Ballard, 1999); Summerfield et al., 2006). However, there has been comparatively little work attempting to explain how, under PP, perceptual information makes the transition from unconscious processing, where it is inaccessible from the first person point of view, to conscious processing. Given that some theorists have argued that PP has the potential to act as a unified theory of the mind/brain (e.g. Clark, 2013; Friston, 2010), this is an especially pertinent issue. One approach to this problem is to try to integrate an independently developed theory of consciousness into the framework of PP. Such an approach was taken by Hohwy (2013), who proposed an account of conscious access that integrates the global neuronal workspace (GNW) – one of the most empirically well supported models of conscious access (Dehaene, 2014; Dehaene & Changeux, 2011) – into the framework of PP. Henceforth, for the sake of brevity, I will refer to Hohwy (2013) account as the predictive global neuronal workspace (PGNW).

E-mail address: [email protected]. https://doi.org/10.1016/j.concog.2019.102763 Received 2 February 2019; Received in revised form 29 April 2019; Accepted 7 June 2019 Available online 19 June 2019 1053-8100/ © 2019 Elsevier Inc. All rights reserved.

Consciousness and Cognition 73 (2019) 102763

C.J. Whyte

When Hohwy proposed the PGNW his primary explanatory target was phenomenal unity and the account has yet to be extended to encompass other phenomena, leaving it both empirically underexplored and theoretically underdeveloped. My aim in this article is to outline the empirical predictions that distinguish the PGNW from other workspace models in the literature. In the following, I briefly review PP and the GNW, and then outline three lines of inquiry that build on and support the PGNW. The first introduces the PGNW, places it in close contact with experimental work, and cashes out two empirical predictions that allow it to be experimentally distinguished from the standard formulation of the GNW. The second explores the evidence for the first of these predictions in the context of bistable perception and the conscious processing of auditory regularities. The third contrasts the PGNW with Chanes and Barrett (2016) limbic workspace – an alternative approach to the integration of a workspace architecture into the framework of predictive processing – and discuss the predictions that differentiate the two accounts. 2. Stage setting part one: Predictive processing If we view the problem of sensory inference from an organism’s perspective (Eliasmith, 2005), it becomes apparent that the only information the brain has about the world is delivered to it via sensory transduction. Brains have no direct access to the causes of these signals and yet we come to perceive a coherent world replete with detail. According to PP, the brain accomplishes this task by generating predictions about the incoming barrage of signals emanating from sensory transducers and then revising the prediction in light of the mismatch between the prediction and the actual signal (this mismatch is termed prediction error). The smaller the prediction error, the better the prediction. Broadly speaking, top-down predictions (priors) are met with bottom-up signals and are revised in light of the prediction error. The newly adjusted prediction serves as the next prior. This process of minimising prediction error is iterated throughout the cortical hierarchy. Bottom-up signals that are not predicted at one level generate prediction errors that are fed forward to the next level in the hierarchy which in turn tries to predict the incoming prediction error signal (Clark, 2016; Friston, 2005; Hohwy, 2013). The content of perception is the prediction that best explains away prediction error throughout this hierarchy and thus achieves the highest posterior probability (Hohwy, 2012). 2.1. Precision-Weighting and attention A key element of PP’s architecture is the functional segregation of error units that pass prediction error forwards and state units that send predictions backwards. Throughout levels of the hierarchy the precision of error units is modulated by both state units at higher levels and lateral connections between error units at the same level (Friston, 2005; 2010). This is significant as the variability and reliability of sensory input and thus the resultant generation of prediction error needs to be context sensitive. This is accomplished via a kind of second-order sensory inference about the precision (inverse of the variance) of the signal that alters the weighting on error units through the modulation of post-synaptic gain (Feldman & Friston, 2010). Using the tools of precision-weighting, PP recasts attention, both exogenous and endogenous, as the modulation of post-synaptic gain. Exogenous attention corresponds to conditions where intense stimuli create a strong signal. Error units throughout the hierarchy have an ‘expectation’ that strong signals are precise and as such have a good signal to noise ratio. Thus, in contexts where there is a strong signal, error units treat the signal as precise leading to the rapid generation of prediction errors greatly increasing the perceptual salience of the attended feature. Endogenous attention, by contrast, is directed by the context dependent top-down modulation of error unit precision (Hohwy, 2012). 2.2. The inferential hierarchy Keeping track of a changing environment requires prediction error minimisation at multiple temporal-scales: from the millisecond to millisecond predictions needed to track the low level changes caused by saccadic eye movements, to the long term ‘expectations’ about social norms that govern interpersonal interactions. According to PP, such spatiotemporal structure is recapitulated throughout the cortex. When prediction errors are generated and fed forward, each successive level in the inferential hierarchy tries to predict the causes of the prediction error signal over increasing spatiotemporal scales (Hohwy, 2013). Within the literature there are two primary accounts of how the inferential hierarchy is implemented1. The first account, advanced by Kiebel, Daunizeau, and Friston (2008), argues that the hierarchy is constructed around an approximately posterior-toanterior axis of prediction. According to this view, primary sensory cortices in more posterior regions minimise prediction error over a rapid temporal scale, and as processing abstracts away from immediate input and moves to more anterior regions of cortex, prediction error is minimised over ever increasing temporal scales. Whilst this account is still in its infancy, it is supported by both anatomical and functional considerations within the visual system and select regions of prefrontal cortex. The account can be extended to other modalities and regions, although at this stage the evidence is highly speculative (for a review see the supplementary material included in Kiebel et al., 2008). As such, we shall focus on evidence from vision and prefrontal cortex. It is well accepted that the visual system processes information in accordance with a posterior-to-anterior spatial hierarchy. Low levels process simple stimulus features such as edges and luminance, and as information ascends the hierarchy processing becomes progressively more abstract. By the mid-level of the ventral stream in lateral occipital cortex (LOC) there are representations of motion and view-point invariant features of rotating objects (Schultz, Chuang, & Vuong, 2007). Similarly, there is evidence that high 1

I am grateful to an anonymous reviewer for insightful comments that brought my attention to the differences between the two models. 2

Consciousness and Cognition 73 (2019) 102763

C.J. Whyte

levels of the dorsal stream in parietal cortex also represent invariant object properties (Konen & Kastner, 2008). As Kiebel et al. (2008) emphasise, this implies that the hierarchy is also temporally organised. To represent view-point or motion invariant features of a stimulus while low level input is rapidly changing, these regions must be integrating information over both space and time. The temporal integration of information also seems crucial for the functional organisation of the lateral prefrontal cortex (PFC). Drawing together data from fMRI and non-human primate electrophysiology, Koechlin and Summerfield (2007) depict the lateral PFC as a hierarchical system of control processes organised around an anterior-posterior axis. Control signals are generated by distinct regions of lateral PFC to select amongst potential actions on the basis of information about the current context and past events. Signals generated on the basis of more temporally distant events are sent from successively more anterior parts of lateral PFC. In line with this, Koechlin, Ody, and Kouneiher (2003) used structural equation modelling of fMRI data to show that the maintenance of rules for action selection was correlated with the coupling of activity between anterior and posterior PFC, while coupling between posterior PFC and premotor cortex was correlated with both the maintenance of rules and current context. Furthermore, in monkeys, lesions to anterior PFC, but not posterior PFC, disrupt the selection of action on the basis of information about past events (Goldman, Rosvold, Vest, & Galkin, 1971; Petrides, 1996). Finally, according to Kiebel et al. (2008) a key region near the top of the predictive hierarchy is the orbitofrontal cortex (OFC), which comes to represent temporally extended and stable states by tracking regularities in the behaviour of subordinate levels. In support of this, the model based fMRI study conducted by Weilnhammer and colleagues found that during an associative learning task, predictions of cue-target reliability learnt over the course of many trials correlated with activity in OFC. Whereas low level predictions about the conditional probabilities of the target stimulus correlated with activity in retinotopic visual cortex (Weilnhammer, Stuke, Sterzer, & Schmack, 2018). The second account, forcefully defended by Barrett and colleagues (Barrett & Simmons, 2015; Chanes & Barrett, 2016), is constructed around the structural model of cortico-cortico connections (Barbas, 1986, 2015; Barbas and Rempel-Clower, 1997). On this view, the direction of information flow in the cortex is determined by the gradient of laminar differentiation. Limbic (agranular and dysgranular) cortices that lack a well differentiated layer IV, sit atop the hierarchy and track long term statistical regularities in the behaviour of lower levels. They send predictions to eulaminate multimodal integration areas that have a more developed layer IV through afferent (feedback) connections that originate in deep layers and terminate in superficial layers. In turn, eulaminate multimodal areas send predictions to unimodal association areas (e.g. extrastriate cortex) through similar feedback connections. Finally, unimodal association areas send predictions to primary sensory cortices that track rapidly evolving changes in input and have the most well developed layer IV (koniocortices). Prediction errors are conveyed in the opposite direction through efferent (feedforward) connections that originate in superficial layers and terminate in deep layers of cortex (Barrett & Simmons, 2015; Chanes & Barrett, 2016). Depending on the region of cortex under analysis the two accounts give similar answers to the question of where in the inferential hierarchy a region is located. For example, the evidence reviewed in this section for a posterior-to-anterior axis of prediction in prefrontal cortex also supports the structural model. Anterior lateral PFC has a less well differentiated layer IV than posterior lateral PFC (eulaminate I and eulaminate II respectively; Chanes & Barrett, 2016; Barbas, 1986, 2015), and OFC has an even simpler laminar structure consisting of eulaminate I and dysgranular cortices (see figure 2 in Barbas, 2015). Thus, these regions occupy similar regions in the inferential hierarchy in both models. The same is also true of the visual system. As such, at this early stage of theorising it is not necessary to commit the PGNW to a particular model of the inferential hierarchy. However, it should be mentioned that the structural model has a clear advantage. Unlike the posterior-anterior-model, the structural model is able to give a definite answer to the question of where in the inferential hierarchy a specific region is located, regardless of sensory modality, as it is determined by the region’s degree of laminar differentiation. With an overview of the relevant aspects of PP now in hand, we now move onto the global neuronal workspace (GNW). 3. Stage setting part Two: The global neuronal workspace Global workspace theory was originally advanced by Baars (1988, 2002, 2005) as a cognitive architecture centred around the hypothesis that consciousness allows diffuse access to information between what would otherwise be relatively isolated processes. Conscious processes have a limited capacity, are serial, widely distributed, flexible, and context sensitive. In contrast, unconscious processes have a much greater informational capacity, function in parallel and are relatively isolated and autonomous. To become conscious, information must enter the global workspace: a serial and widely distributed means of information exchange that facilitates interaction between numerous otherwise isolated functions. Baars’ (1988, 1997) presentation of global workspace theory was constructed around an experimental approach to studying consciousness termed contrastive phenomenology. Baars argued that by using paradigms that allow stimulus visibility to be manipulated (e.g. by masking, inattention or binocular rivalry, etc.), whilst holding all other variables as constant as possible, consciousness could be treated as an independent variable, allowing inferences to be made about the computational roles of conscious and unconscious information processing. Using the contrastive approach in combination with neuroimaging, Dehaene and colleagues have advanced the GNW, an account of the global workspace’s neurobiological implementation (Dehaene, 2014; Dehaene & Changeux, 2011; Dehaene, Changeux, & Naccache, 2011; Dehaene & Naccache, 2001). Since the primary source of evidence for the GNW comes from paradigms that operationalise consciousness as reportability, this paper will be chiefly be concerned with what Block (2005) calls “access-consciousness”, which is defined as the availability of information for control of action, executive processes, and verbal report. 3

Consciousness and Cognition 73 (2019) 102763

C.J. Whyte

3.1. Implementation of the workspace According to Dehaene and colleagues, the global workspace is implemented in a network of densely connected pyramidal neurons possessing long-range excitatory axons connecting prefrontal2 and parietal cortices. Information that is integrated into the global workspace is broadcast and maintained by re-entrant activity between a fragment of the workspace neurons, while the majority of neurons are inhibited. The wide scale inhibition of competing input processes transiently prevents information from other sources from being broadcast, giving rise to serial information processing. Broadcasting to this system facilitates more efficient and flexible computation as numerous otherwise informationally encapsulated subsystems are able to process and share the same piece of information seconds after its initial presentation. Thus, encoding information in the global workspace enables numerous high level cognitive functions including planning, verbal report, conscious working memory and voluntary motor behaviour (Dehaene & Changeux, 2011; Dehaene et al., 2011). A central theoretical element of this account is the prediction that the initial activation of sensory regions to invisible stimuli under conditions of masking or inattention will be almost identical to visible stimuli (Dehaene, Changeux, Naccache, Sackur, & Sergent, 2006). In contrast, during the relatively late time window of ∼250–500 ms post stimulus onset, visible stimuli are predicted to result in the non-linear and all-or-nothing ignition of a long-ranging prefrontal-parietal network (Dehaene & Changeux, 2011). In neural terms, recurrently connected neurons compete in parallel in a process of evidence accumulation whereby they either cross an evidential threshold and have their information integrated into the global workspace, or they remain unconscious (Dehaene, 2011; Dehaene, Charles, King, & Marti, 2014). If ignition fails to occur, unconscious representations are used to guide behaviour (Del Cul, Dehaene, Reyes, Bravo, & Slachevsky, 2009). This prediction is well supported experimentally. Using a metacontrast masking paradigm in EEG, Del Cul and colleagues showed that early ERP components (P1 and N1) do not display a significant difference between seen and unseen trials. The relatively late P300 (P300b) component, by contrast, displays a significant non-linear increase in amplitude between seen and unseen trials (Del Cul, Baillet, & Dehaene, 2007). The P300 is interpreted by Dehaene (2014) as an index of conscious processing. The integration of information into the global workspace involves the depolarisation of only a fraction of the workspace neurons while the majority are inhibited, explaining the positivity over central electrodes that characterises the P300. Similarly, in fMRI Dehaene and colleagues showed that under conditions of masking, the BOLD signal remains within the left fusiform gyrus, extrastriate cortex and left precentral sulcus. Whereas, when a stimulus is unmasked there is wide spread activation of prefrontal, temporal and parietal regions (Dehaene et al., 2001, p. 753). In summary then, according to the GNW conscious access is characterised by the wide-spread, late, and all-or-nothing activation of the prefrontal and parietal network that is the global workspace. With a synoptic view of PP and the GNW now in hand we are in a position to explore the integration of the two accounts, the task to which I now turn. 4. Prediction and the workspace As noted in the introduction, many theorists tout PP’s ability to unify seemingly disparate processes (Friston, 2005, 2010; Lee & Mumford, 2003; Rao & Ballard, 1999; Richter, Ekman, & de Lange, 2018; Summerfield et al., 2006). There is, however, relatively little discussion of the difference between conscious and unconscious processing within the PP literature. The GNW was developed specifically to account for this difference. As such, Hohwy (2013) effort to combine these two frameworks with the PGNW is worthy of further investigation. The following sections advance three lines of argument that build on the PGNW, beginning with an outline of the theory and the empirical predictions that distinguish it from the standard model of the GNW. 4.1. The predictive global neuronal Workspace. As we have seen during ignition, recurrently connected neurons compete in parallel in a process of evidence accumulation, whereby they either cross an evidential threshold and are integrated into the prefrontal-parietal network that is the global workspace, or they remain unconscious. The small fraction of neurons that achieve ignition by crossing this evidential threshold broadcasts information to consuming subsystems through recurrent activity while the majority are transiently inhibited (Dehaene & Changeux, 2011). Since there is considerable evidence that conscious access depends upon ignition, the critical question is, what is it that explains the threshold of ignition? Hohwy (2013) has proposed that by translating the process of evidence accumulation into the framework of PP, a possible answer emerges. According to PP, what we come to perceive is the prediction that best explains away prediction error throughout the hierarchy and consequently achieves the highest posterior probability (Hohwy, 2012). This is of central importance to the guidance of action (active inference within the framework of PP). In order for the brain to select the action policy that will best minimise prediction error given its current model of the world, the brain must, at least transiently, hold this model fixed. Therefore, the process of ignition – whereby the most probable representation of the world is selected and broadcast – is explained by the need to switch from perceptual to active inference. This gives a principled reason for the existence of a threshold, it represents the point at which a prediction about the structure of the world has sufficient evidence to warrant guiding action. Thus, according to the PGNW, 2 There is considerable debate about whether frontal cortex is a genuine neural correlate of consciousness or merely reflects verbal report (Boly et al., 2017; Odegaard, Knight & Lau, 2017). Although, I flag this issue only to put it aside as it is outside the purview of this article.

4

Consciousness and Cognition 73 (2019) 102763

C.J. Whyte

ignition occurs when a prediction is better able to minimise prediction error than competing predictions and is thereby the most probable explanation of current sensory input. Since the influence of prediction versus prediction error is affected by uncertainty, uncertainty will inevitably modulate the threshold for ignition. This formulation of ignition entails that top-down connections from the global workspace to input processes carry predictions that play an actively constructive role in inferring the contents of consciousness. Although admitting of multiple interpretations, consistent with the above proposal, there is evidence that feedback from global workspace regions to sensory cortices plays a central role in conscious access. Specifically, Gaillard et al. (2009) used Granger causal analysis to measure the influence between intracranial electrodes implanted in epilepsy patients while they were presented with masked and unmasked words. In essence, Granger causality works by comparing how well two autoregressive models predict time series data. To take a toy example, if we consider two variables X and Y and are trying to measure the direction of influence between X and Y, X is said to G-cause Y, if X contains information that helps predict Y better than the past information of Y alone (Friston, Moran, & Seth, 2013; Seth, 2007; Seth, Barrett, & Barnett, 2015). Consistent with the PGNW, in both unmasked and masked conditions, there was bidirectional influence between posterior and anterior electrode pairs between frontal and sensory cortices, indicating an interaction between top-down predictions (anterior-toposterior influence) and bottom-up prediction error (posterior-to-anterior influence) during the time window of ignition (Gaillard et al., 2009, p. 481–485). Similarly, dynamic causal modelling of the ERPs elicited during a mismatch negativity paradigm shows that, compared to healthy controls, patients in a vegetative state have preserved feed-forward connections but impaired feedback connections from frontal cortex to superior temporal cortex (Boly et al., 2011). And, dynamic causal modelling of the BOLD response elicited by perceptual transitions in a bistable perception paradigm has shown that perceptual transitions are associated with increased top-down connectivity from frontal to visual cortices (Weilnhammer, Ludwig, Hesselmann, & Sterzer, 2013). As I alluded to above, while this evidence is suggestive of prediction error minimisation it can also be interpreted as supporting the standard model of the GNW. Feedback from frontal to sensory cortices is not synonymous with prediction error minimisation. Crucially, however, there are at least two hypotheses that are unique to the PGNW. The first hypothesis concerns the temporal dynamics of conscious representation. Specifically, if the global workspace is underwritten by a predictive coding architecture, conscious processing should be continuous with the rest of the inferential hierarchy. This is not to say that there are not important differences in processing at different levels of the hierarchy3. Processing at high levels of the hierarchy may, for example, involve discrete representation while processing at lower levels may be continuous (think of categorical representations versus representations of a quantity like luminance; Friston, Parr, and de Vries (2017). I simply mean that every level of the hierarchy will be involved in the process of prediction error minimisation via hierarchical message passing. Notably, this entails that the stability of conscious representation should be sensitive to the accumulation of prediction error at lower levels of the hierarchy. This was initially proposed by Hohwy et al. (2008) in the context of binocular rivalry, and has subsequently become a subject of debate in the context of the conscious processing of auditory regularities (King, Gramfort, Schurger, Naccache, & Dehaene, 2014; Shirazibeheshti et al., 2018). We turn to the details of this hypothesis in the following section. The second hypothesis concerns the role of uncertainty in the modulation of top-down prediction versus feed-forward prediction error. Recall that according to PP each successive level of the inferential hierarchy tries to predict the causes of the prediction error signal over increasing spatiotemporal scales. Both the structural model (Barrett & Simmons, 2015; Chanes & Barrett, 2016) and the posterior-to-anterior model (Kiebel et al., 2008) of the inferential hierarchy place the prefrontal-parietal network that constitutes the global workspace at a higher level of the inferential hierarchy than the input networks in sensory areas. Thus, in experimental settings where prior uncertainty is low, predictions sent via feedback connections from the global workspace to input networks will attenuate prediction error. And when uncertainty is high, prediction errors sent in the opposite direction will have greater influence. In the context of ignition, this implies that when priors are precise, top-down predictions will have more influence on the content of consciousness, whereas when the bottom-up signal is judged to be precise, or when predictions are violated, the prediction error signal will have more influence. This can be tested in M/EEG by operationalising the influence of prediction and prediction error in terms of the bias in Granger causality4 between anterior (frontal) and posterior (occipital) sensors (for an example of a similar method see Goddard, Carlson, Dermody, & Woolgar, 2016), in combination with a paradigm that allows for the manipulation of perceptual uncertainty and stimulus visibility5 (given the possibility of correlation between nearby sensors/electrodes this method is best suited to visual experiments). If the PGNW is on track, during the time window of ignition (300–500 ms post stimulus onset), when uncertainty is low (i.e. priors are precise) there will be a greater anterior-posterior bias between sensors. In turn, when uncertainty is high, or predictions are violated, there will be a greater posterior-anterior bias. Importantly, this hypothesis is distinct from the model of standard model of the GNW defended by Dehaene and Changeux (2011), which implies that conscious access should always be characterised by a feedforward sweep of information as the winning representation gets fed forward from sensory areas into the workspace (see Gaillard et al., 2009, p. 481–485 for an explicit example in an experimental context). Overall then, the PGNW is consistent with existing evidence showing that feedback from prefrontal-parietal regions to sensory cortices plays a crucial role in conscious access. And further, it generates at least two surplus hypotheses that allow it to be distinguished from the standard formulation of the GNW. The next section examines empirical evidence in support of the first hypothesis.

3

The clarity of this point was greatly improved by helpful comments from an anonymous reviewer. This hypothesis is also readily testable via dynamic causal modelling. 5 Since attention reverses the silencing effect of prediction (Kok et al., 2011) it will also need to be controlled for. 4

5

Consciousness and Cognition 73 (2019) 102763

C.J. Whyte

4.2. Prediction error minimisation and the stability of conscious Representation. According to the PGNW, the architecture that underwrites the global workspace is continuous with the rest of the inferential hierarchy in the sense that, like the rest of the hierarchy, the global workspace is engaged in a process of hierarchical prediction error minimisation. This entails that the stability of conscious representation will be modulated by the accumulation of prediction error sent from lower levels of the hierarchy. This hypothesis was first advanced by Hohwy et al. (2008) as an explanation of alternation in binocular rivalry. According to their account no stimulus has both a high likelihood and prior so when one stimulus dominates during rivalry prediction error for the suppressed stimulus accumulates, eventually causing a perceptual transition. This has subsequently been tested by Weilnhammer et al. (2017) using model based fMRI. In line with Hohwy et al. (2008) hypothesis, the parametric modulator6 encoding the trajectory of the prediction error signal was predictive of perceptual transitions and correlated significantly with the BOLD signal in bilateral inferior frontal gyri and bilateral insula (both regions implicated in the PGNW). Notice that this would not be predicted by a purely feed-forward account as prediction error is only predictive of perceptual transitions because it leads to an update of the dominant percept’s prior. Speculatively, this may also help to explain why rivalry is slowed in the absence of attention (Paffen, Alais, & Verstraten, 2006). Since attention increases the precision of prediction error, in the absence of attention prediction errors will carry less weight and take longer to accumulate the level of evidence required to update the dominant precepts prior (Hohwy et al., 2008). Conversely, in the context of the conscious processing of auditory regularities, this hypothesis has been challenged by King et al. (2014). Specifically, in the local-global paradigm (Bekinschtein et al., 2009), which manipulates sensory expectations over two distinct time scales. The paradigm presents participants with a sequence of five tones, the first four of which have an identical frequency, while the fifth tone either confirms the predicted pattern (local standard), or deviates from it (local deviation) by presenting a higher or lower frequency tone eliciting the MMN component. The frequency of the local standards and local deviations on the last trail of each sequence establishes a global pattern that can also be violated (global deviation) eliciting the P300 component. King et al. (2014) argue that the MMN component elicited by local violations is an index of unconscious prediction error, whereas the P300 is an index of a separate conscious working memory system updating and maintaining information through to the next trail. Accordingly, unconscious prediction error minimisation and the updating of working memory are argued to predict qualitatively distinct neural dynamics. When prediction errors are generated they are rapidly propagated forward until they are matched by topdown predictions, leading to a highly dynamic pattern of activation. In contrast, the active maintenance of information in working memory creates a stable pattern of activation. King and colleagues tested this via the time generalisation extension of multivariate pattern analysis (Grootswagers, Wardle, & Carlson, 2017; King & Dehaene, 2014). They predicted that classifiers trained to discriminate between local deviations and local standards would not generalise across time as the pattern of activity changes dynamically as prediction error is fed forward, whilst classifiers trained to discriminate between global standards and global violations would generalise across time as the pattern of activity maintained by working memory would remain stable. This is precisely what was found. Classifiers trained to discriminate between local standards and local deviations did not generalise across time, while classifiers trained to discriminate between global standards and global violations reliably generalised between 125 and 700 ms post stimulus onset (King et al., 2014, p. 4). At first glance this seems problematic for PGNW. However, contrary to King et al. (2014) the stable maintenance of a distinct neural pattern is not necessarily opposed to a prediction error minimisation architecture. When prediction errors are generated and fed forward, each successive level in the inferential hierarchy predicts the causes of the prediction error signal at increasing levels of spatiotemporal abstraction (Hohwy, 2013). The prefrontal-parietal network implicated in the GNW is considerably higher in the inferential hierarchy than the sensory areas that support the response to local deviations. Therefore, it is not surprising that the prediction error response to the global deviation is sustained for a longer period of time. As Shirazibeheshti et al. (2018) argue, the crucial question is whether the stability of the response evoked by the global deviation can be modulated by the accumulation of prediction error at lower levels. If the computations that support the global workspace are continuous with the inferential hierarchy, then the accumulation of prediction error at lower levels of the hierarchy should both accelerate and shorten the P300 response, whereas, if the global workspace employs a mode of computation that is distinct from the inferential hierarchy, as King et al. (2014) suggest, the amount of prediction error at lower levels should not affect the latency of the P300 component. Shirazibeheshti et al. (2018) tested this by factorally manipulating local and global deviations. In line with the predictions of the PGNW, they found that when global deviations were preceded by local deviations the P300 response was accelerated (i.e. it occurred earlier) and was also attenuated more rapidly. Taken together the findings support the hypothesis that conscious processing is continuous with the rest of the inferential hierarchy, contrary to the challenge put forward by King et al. (2014). Having explored the evidence for the PGNW, the next section compares the PGNW to an alternative approach to the integration of a workspace architecture into the framework of PP proposed by Chanes and Barrett (2016). 4.3. The limbic workspace versus the predictive global neuronal workspace Hohwy (2013) is not the only researcher to argue for an account of conscious access that integrates a workspace architecture into PP. A similar but ultimately distinct account has been advanced by Chanes and Barrett (2016) who explicitly locate their view within the structural model of the inferential hierarchy (see also Barrett & Simmons, 2015). 6

Parametric modulators test for a linear relationship between the BOLD signal and some variable of interest (e.g. reaction time). 6

Consciousness and Cognition 73 (2019) 102763

C.J. Whyte

Unlike the PGNW, which following the GNW identifies the global workspace with lateral prefrontal and parietal areas primarily consisting of eulaminate multimodal integration areas, Chanes and Barrett (2016) propose that the limbic cortices that sit atop the hierarchy form a “limbic workspace” that supports conscious access. Limbic cortices have dense bidirectional connections with subordinate levels of the hierarchy, and subcortical structures such as the hypothalamus and the amygdala. As such, Chanes and Barrett suggest that limbic cortices (cingulate cortex, anterior insula, posterior OFC, parahippocampal gyrus and the temporal pole) select between unimodal representations, or combinations thereof, on the basis of their relevance to the organism’s long term homeostatic needs and preferences, and integrate them into a unified conscious representation. Limbic cortices also play important functional roles in the PGNW and standard model of the GNW (Dehaene & Naccache, 2001), however, neither of these structures are argued to be necessary for conscious access. The anterior insula (aI) is hypothesised to play a role in the selection of content for consciousness by prioritising the allocation of attentional resources across modalities (Michel, 2017). And, the anterior cingulate cortex (ACC) is heavily implicated in conscious conflict monitoring (Dehaene, 2018; Dehaene et al., 2003). Thus, to borrow Aru, Bachmann, Singer, and Melloni (2012) nomenclature, the aI’s hypothesised role is best described as a prerequisite to consciousness, and the ACC’s role is best described as a consequence of consciousness. In contrast, lateral PFC and posterior parietal cortex, both eulaminate in structure (Barbas, 2015; Medalla & Barbas, 2006), are argued to play necessary causal roles in conscious access (Del Cul et al., 2009). Hence, the key difference between the PGNW and the limbic workspace is that the limbic workspace attributes an integrative role to limbic cortices that is crucial for conscious access and is not just a prerequisite to, or consequences of, conscious access. Since both the PGNW and the limbic workspace are underwritten by a predictive coding architecture the hypotheses set out in Sections 4.1 and 4.2 will not distinguish between the two models. Instead, testing between them will require systematic inter-level experiments (Craver, 2007) involving both “top-down” neuroimaging experiments investigating whether activity in limbic cortices distinguishes between conscious versus unconscious processing across sensory modalities, and “bottom-up” interventions that investigate whether lesioning (either artificially via TMS or by testing patient populations) limbic cortices disrupts conscious access. For example, in the context of visual consciousness a meta-analysis of bistable perception and backwards masking studies found a network of regions in extrastriate, temporal, parietal and prefrontal cortices that was active in conscious but not unconscious conditions (Bisenius, Trapp, Neumann, & Schroeter, 2015). Importantly, apart from the anterior insula (aI) no other limbic region was found to reliably correlate with visual consciousness across paradigms (Bisenius et al., 2015). If combined with a bottom-up experiment showing that patients with lesions to the aI do not have impairments to visual consciousness this would be a strike against the limbic workspace. At least for visual consciousness. As it stands, however, this kind of bottom-up evidence only exists for lateral prefrontal and parietal cortices7. Specifically, Del Cul et al. (2009) found that patients with lateral prefrontal damage, in comparison to healthy controls, have a higher backwards masking threshold even while controlling for attentional deficits. In addition, thetaburst TMS to bilateral dlPFC has been shown to impair metacognitive visual awareness (Rounis, Maniscalco, Rothwell, Passingham, & Lau, 2010). Repetitive TMS to ventral parietal cortex has been found to significantly decrease the percentage of trails in which subjects report consciously perceiving a cue that is presented at threshold (Babiloni et al., 2006). And, TMS to the intraparietal sulcus has been shown to induce perceptual fading (Kanai, Muggleton, & Walsh, 2008). Unfortunately, because limbic cortices are often hidden beneath other layers of cortex, non-invasive stimulation studies are not possible. Instead we must rely on patients with lesions to limbic cortices which are rare, although not unheard of (Damasio, Damasio & Tranel, 2013; Philippi et al., 2012). A good illustration of the type of work that is needed comes from the patient tested by Khalsa, Rudrauf, Feinstein, and Tranel (2009) who had almost complete bilateral lesions to the aI and ACC. Interestingly, the lesions seemed cause a delay in interoceptive awareness of heartbeat sensations in comparison to healthy controls. Although it must be emphasised that single patient studies like this must be interpreted with caution, diffuse damage to surrounding areas and differences in the location of lesions between patients can introduce confounds. In sum, at present there is insufficient evidence to distinguish between the two accounts, however, this is not to say that it is not possible. Testing between the PGNW and the limbic workspace will require systematic examination of conscious access across sensory modalities in patients with lesions to limbic cortices, in addition to meta-analyses of neuroimaging experiments investigating conscious access across sensory modalities, and experimental paradigms, of the kind conducted by Bisenius et al. (2015) in the context of vision. 5. Concluding remarks Recent years have seen an explosion of interest in PP across a wide range of areas within cognitive science. Still, there remains a gap in the literature when it comes to explaining conscious access in terms of the computational principles that underwrite prediction error minimisation. The aim of this paper has been to begin the process of closing this gap by advancing three lines of inquiry aimed at distinguishing the PGNW from other workspace models in the literature. The first demonstrated that as well as being consistent with current evidence, the PGNW generates surplus hypotheses that distinguish it from the standard model of the GNW. The second reviewed evidence from bistable perception and the conscious processing of auditory regularities in favour of the hypothesis that conscious representation is sensitive to prediction error at lower levels of the inferential hierarchy. The third compared the PGNW to Chanes and Barrett (2016) limbic workspace and outlined a set of predictions that will allow the two accounts to be teased apart. As it stands, therefore, the PGNW is well positioned to guide both future empirical and theoretical research on the role of prediction error minimisation in conscious access. 7

For dissenting opinion about the involvement of parietal cortex see Lau & Passingham, 2006; Lau & Rosenthal, 2011. 7

Consciousness and Cognition 73 (2019) 102763

C.J. Whyte

Acknowledgments The author would like to thank Elias Dokos, Marianne McAllister, Alexander Gillett, Stephen Gadsby, Christopher Hewitson, Michael Kirchoff, Inger Southwick, Thomas Carlson, David Kaplan and two anonymous reviewers for invaluable feedback and discussion on the topic of this article. References Aru, J., Bachmann, T., Singer, W., & Melloni, L. (2012). Distilling the neural correlates of consciousness. Neuroscience & Biobehavioral Reviews, 36(2), 737–746. Baars, B. J. (1988). A cognitive theory of consciousness. Cambridge, UK: Cambridge University Press. Baars, B. J. (1997). Contrastive Phenomenology. In N. J. Block, O. J. Flanagan, & G. Güzeldere (Eds.). The nature of consciousness: Philosophical debates (pp. 187–201). MIT press. Baars, B. J. (2002). The conscious access hypothesis: Origins and recent evidence. Trends in Cognitive Sciences, 6(1), 47–52. Baars, B. J. (2005). Global workspace theory of consciousness: Toward a cognitive neuroscience of human experience. Progress in Brain Research, 150, 45–53. Babiloni, C., Vecchio, F., Rossi, S., De Capua, A., Bartalini, S., Ulivelli, M., et al. (2006). Human ventral parietal cortex plays a functional role on visuospatial attention and primary consciousness. A repetitive transcranial magnetic stimulation study. Cerebral Cortex, 17, 1486–1492. Barrett, L. F., & Simmons, W. K. (2015). Interoceptive predictions in the brain. Nature Reviews Neuroscience, 16(7), 419. Barbas, H. (1986). Pattern in the laminar origin of corticocortical connections. Journal of Comparative Neurology, 252(3), 415–422. Barbas, H. (2015). General cortical and special prefrontal connections: Principles from structure to function. Annual Review of Neuroscience, 38, 269–289. Barbas, H (1997). Cortical structure predicts the pattern of corticocortical connections. Cerebral cortex (New York, NY: 1991), 7(7), 635–646. https://doi.org/10.1093/ cercor/7.7.635. Bekinschtein, T. A., Dehaene, S., Rohaut, B., Tadel, F., Cohen, L., & Naccache, L. (2009). Neural signature of the conscious processing of auditory regularities. Proceedings of the National Academy of Sciences, 106(5), 1672–1677. Bisenius, S., Trapp, S., Neumann, J., & Schroeter, M. L. (2015). Identifying neural correlates of visual consciousness with ALE meta-analyses. Neuroimage, 122, 177–187. Block, N. (2005). Two neural correlates of consciousness. Trends in Cognitive Sciences, 9(2), 46–52. Boly, M., Garrido, M. I., Gosseries, O., Bruno, M. A., Boveroux, P., Schnakers, C., et al. (2011). Preserved feedforward but impaired top-down processes in the vegetative state. Science, 332(6031), 858–862. Boly, M., Massimini, M., Tsuchiya, N., Postle, B. R., Koch, C., & Tononi, G. (2017). Are the neural correlates of consciousness in the front or in the back of the cerebral cortex? Clinical and neuroimaging evidence. Journal of Neuroscience, 37(40), 9603–9613. Chanes, L., & Barrett, L. F. (2016). Redefining the role of limbic areas in cortical processing. Trends in Cognitive Sciences, 20(2), 96–106. Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(03), 181–204. Clark, A. (2016). Surfing uncertainty: Prediction, action, and the embodied mind. Oxford University Press. Craver, C. F. (2007). Explaining the brain: Mechanisms and the mosaic unity of neuroscience. Oxford University Press. Damasio, A., Damasio, H., & Tranel, D. (2012). Persistence of feelings and sentience after bilateral damage of the insula. Cerebral Cortex, 23(4), 833–846. Dehaene, S. (2011). Conscious and nonconscious processes: Distinct forms of evidence accumulation?. In Biological physics. Basel: Springer141–168. Dehaene, S. (2014). Consciousness and the brain: Deciphering how the brain codes ourthoughts. Penguin. Dehaene, S. (2018). The error-related negativity, self-monitoring, and consciousness. Perspectives on Psychological Science, 13(2), 161–165. Dehaene, S., Artiges, E., Naccache, L., Martelli, C., Viard, A., Schürhoff, F., et al. (2003). Conscious and subliminal conflicts in normal subjects and patients with schizophrenia: The role of the anterior cingulate. Proceedings of the National Academy of Sciences, 100(23), 13722–13727. Dehaene, S., Changeux, J. P., & Naccache, L. (2011). The global neuronal workspace model of conscious access: From neuronal architectures. to clinical applications. In Characterizing consciousness: From cognition to the clinic? (pp. 55–84). Heidelberg: Springer, Berlin. Dehaene, S., & Changeux, J. P. (2011). Experimental and theoretical approaches to conscious processing. Neuron, 70(2), 200–227. Dehaene, S., Charles, L., King, J. R., & Marti, S. (2014). Toward a computational theory of conscious processing. Current opinion in neurobiology, 25, 76–84. Dehaene, S., Changeux, J. P., Naccache, L., Sackur, J., & Sergent, C. (2006). Conscious, preconscious, and subliminal processing: A testable taxonomy. Trends in Cognitive Sciences, 10(5), 204–211. Dehaene, S., & Naccache, L. (2001). Towards a cognitive neuroscience of consciousness: Basic evidence and a workspace framework. Cognition, 79(1), 1–37. Dehaene, S., Naccache, L., Cohen, L., Le Bihan, D., Mangin, J. F., Poline, J. B., et al. (2001). Cerebral mechanisms of word masking and unconscious repetition priming. Nature Neuroscience, 4(7), 752–758. Del Cul, A., Baillet, S., & Dehaene, S. (2007). Brain dynamics underlying the nonlinear threshold for access to consciousness. PLoS Biology, 5(10), e260. Del Cul, A., Dehaene, S., Reyes, P., Bravo, E., & Slachevsky, A. (2009). Causal role of prefrontal cortex in the threshold for access to consciousness. Brain, 132(9), 2531–2540. Eliasmith, C. (2005). A new perspective on representational problems. Journal of Cognitive Science, 6(97), 123. Feldman, H., & Friston, K. (2010). Attention, uncertainty, and free-energy. Frontiers in Human Neuroscience, 4, 215. Fletcher, P. C., & Frith, C. D. (2009). Perceiving is believing: A Bayesian approach to Explaining the positive symptoms of schizophrenia. Nature Reviews Neuroscience, 10(1), 48–58. Friston, K. (2005). A theory of cortical responses. Philosophical Transactions of the Royal Society of London B: Biological Sciences, 360(1456), 815–836. Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. Friston, K. J., Parr, T., & de Vries, B. (2017). The graphical brain: Belief propagation and active inference. Network Neuroscience, 1(4), 381–414. Friston, K., Moran, R., & Seth, A. K. (2013). Analysing connectivity with Granger causality and dynamic causal modelling. Current Opinion in Neurobiology, 23(2), 172–178. Gaillard, R., Dehaene, S., Adam, C., Clémenceau, S., Hasboun, D., Baulac, M., et al. (2009). Converging intracranial markers of conscious access. PLoS Biology, 7(3), e1000061. Goddard, E., Carlson, T. A., Dermody, N., & Woolgar, A. (2016). Representational dynamics of object recognition: Feedforward and feedback information flows. Neuroimage, 128, 385–397. Goldman, P. S., Rosvold, H. E., Vest, B., & Galkin, T. W. (1971). Analysis of the delayed-alternation deficit produced by dorsolateral prefrontal lesions in the rhesus monkey. Journal of Comparative and Physiological Psychology, 77(2), 212. Grootswagers, T., Wardle, S. G., & Carlson, T. A. (2017). Decoding dynamic brain patterns from evoked responses: A tutorial on multivariate pattern analysis applied to time series neuroimaging data. Journal of Cognitive Neuroscience, 29(4), 677–697. Hohwy, Jakob (2012). Attention and conscious perception in the hypothesis testing brain. Frontiers in Psychology, 3. https://doi.org/10.3389/fpsyg.2012.00096. Hohwy, J. (2013). The predictive mind. Oxford University Press. Hohwy, J., Roepstorff, A., & Friston, K. (2008). Predictive coding explains binocular rivalry: An epistemological review. Cognition, 108(3), 687–701. Kanai, R., Muggleton, N. G., & Walsh, V. (2008). TMS over the intraparietal sulcus induces perceptual fading. Journal of Neurophysiology, 100(6), 3343–3350. Khalsa, S. S., Rudrauf, D., Feinstein, J. S., & Tranel, D. (2009). The pathways of interoceptive awareness. Nature Neuroscience, 12(12), 1494. King, J. R., & Dehaene, S. (2014). Characterizing the dynamics of mental representations: The temporal generalization method. Trends in Cognitive Sciences, 18(4), 203–210. King, J. R., Gramfort, A., Schurger, A., Naccache, L., & Dehaene, S. (2014). Two distinct dynamic modes subtend the detection of unexpected sounds. PloS One, 9(1),

8

Consciousness and Cognition 73 (2019) 102763

C.J. Whyte

e85791. Kiebel, S. J., Daunizeau, J., & Friston, K. J. (2008). A hierarchy of time-scales and the brain. PLoS Computational Biology, 4(11), e1000209. Koechlin, E., Ody, C., & Kouneiher, F. (2003). The architecture of cognitive control in the human prefrontal cortex. Science, 302(5648), 1181–1185. Koechlin, E., & Summerfield, C. (2007). An information theoretical approach to prefrontal executive function. Trends in Cognitive Sciences, 11(6), 229–235. Kok, P., Rahnev, D., Jehee, J. F., Lau, H. C., & de Lange, F. P. (2011). Attention reverses the effect of prediction in silencing sensory signals. Cerebral cortex, bhr310. Konen, C. S., & Kastner, S. (2008). Two hierarchically organized neural systems for object information in human visual cortex. Nature Neuroscience, 11(2), 224. Lau, H. C., & Passingham, R. E. (2006). Relative blindsight in normal observers and the neural correlate of visual consciousness. Proceedings of the National Academy of Sciences, 103(49), 18763–18768. Lau, H., & Rosenthal, D. (2011). Empirical support for higher-order theories of conscious awareness. Trends in cognitive sciences, 15(8), 365–373. Lee, T. S., & Mumford, D. (2003). Hierarchical Bayesian inference in the visual cortex. JOSA A, 20(7), 1434–1448. Medalla, M., & Barbas, H. (2006). Diversity of laminar connections linking periarcuate and lateral intraparietal areas depends on cortical structure. European Journal of Neuroscience, 23(1), 161–179. Michel, M. (2017). A role for the anterior insular cortex in the global neuronal workspace model of consciousness. Consciousness and Cognition, 49, 333–346. Odegaard, B., Knight, R. T., & Lau, H. (2017). Should a few null findings falsify prefrontal theories of conscious perception? Journal of Neuroscience, 37(40), 9593–9602. Paffen, C. L., Alais, D., & Verstraten, F. A. (2006). Attention speeds binocular rivalry. Psychological Science, 17(9), 752–756. Petrides, M. (1996). Specialized systems for the processing of mnemonic information within the primate frontal cortex. Philosophical Transactions of the Royal Society B, 351(1346), 1455–1462. Philippi, C. L., Feinstein, J. S., Khalsa, S. S., Damasio, A., Tranel, D., Landini, G., et al. (2012). Preserved self-awareness following extensive bilateral brain damage to the insula, anterior cingulate, and medial prefrontal cortices. PloS One, 7(8), e38413. Rao, R. P., & Ballard, D. H. (1999). Predictive coding in the visual cortex: A functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1), 79–87. Richter, D., Ekman, M., & de Lange, F. P. (2018). Suppressed sensory response to predictable object stimuli throughout the ventral visual stream. Journal of Neuroscience, 3421–3517. Rounis, E., Maniscalco, B., Rothwell, J. C., Passingham, R. E., & Lau, H. (2010). Theta-burst transcranial magnetic stimulation to the prefrontal cortex impairs metacognitive visual awareness. Cognitive Neuroscience, 1(3), 165–175. Schultz, J., Chuang, L., & Vuong, Q. C. (2007). A dynamic object-processing network: Metric shape discrimination of dynamic objects by activation of occipitotemporal, parietal, and frontal cortices. Cerebral Cortex, 18(6), 1302–1313. Seth, A. (2007). Granger causality. Scholarpedia, 2(7), 1667.824. Seth, A. K., Barrett, A. B., & Barnett, L. (2015). Granger causality analysis in neuroscience and neuroimaging. Journal of Neuroscience, 35(8), 3293–3297. Shirazibeheshti, A., Cooke, J., Chennu, S., Adapa, R., Menon, D. K., Hojjatoleslami, S. A., et al. (2018). Placing meta-stable states of consciousness within the predictive coding hierarchy: The deceleration of the accelerated prediction error. Consciousness and Cognition, 63, 123–142. Summerfield, C., Egner, T., Greene, M., Koechlin, E., Mangels, J., & Hirsch, J. (2006). Predictive codes for forthcoming perception in the frontal cortex. Science, 314(5803), 1311–1314. Weilnhammer, V., Stuke, H., Hesselmann, G., Sterzer, P., & Schmack, K. (2017). A predictive coding account of bistable perception - a model-based fMRI study. PLoS Computational Biology, 13(5), e1005536. Weilnhammer, V., Stuke, H., Sterzer, P., & Schmack, K. (2018). The neural correlates of hierarchical predictions for perceptual decisions. Journal of Neuroscience, 2901–2917. Weilnhammer, V. A., Ludwig, K., Hesselmann, G., & Sterzer, P. (2013). Frontoparietal cortex mediates perceptual transitions in bistable perception. Journal of Neuroscience, 33(40), 16009–16015.

9