3 Is Our Brain an Open or Closed System? Prediction Model of Brain and World–Brain Relation
Introduction
General Background
How can we be confident in rejecting the possibility of a Cartesian demon that is working to ensure that we are deceived about the real nature of objects and events in the world? If our mind is an open system and thus closely coupled with the world, possible deception about the objects and events within that very same world is unlikely. If, in contrast, our mind remains a closed or self-evidencing system that is inferentially secluded from the world, the door to skepticism and hence for a possible Cartesian demon must be left ajar (see, e.g., Hohwy, 2007, 2013, 2014). Accordingly, the question of whether to characterize the mind as an open or closed system has major epistemic implications.
As do many other philosophers, I believe that mind and its features have their basis in the brain and its neural operations. Although this position is not without opposition, this chapter is not an occasion to adjudicate the metaphysical controversy concerning the relation between the mind and the brain. Here, I am exploring the epistemic implications of a widespread and important conception of how the brain interacts with the world: predictive coding. Given how many neuroscientists are working on predictive coding, and the way that philosophical appeals to neuroscience are increasingly common, thoroughly exploring the epistemic implications of predictive coding should be useful to several scholarly communities. This chapter is about how the brain and its neural operations are related to the world—that amounts to what I conceptually describe as the “world–brain relation.”
Recent work in neuroscience seems to indicate that predictive coding is the main and overarching informational strategy of the brain (Friston, 2010). Predictive coding argues that the neural activity we observe in the brain in response to specific stimuli or tasks does not exclusively result from the stimulus or input alone but, rather from the comparison between the actual input and a predicted input. The brain itself generates a prediction or anticipation of an input, which is then matched and compared with the actual input. The difference between these is called the prediction error, which is taken to be the primary determinant of stimulus-induced activity in the brain.
Predictive coding is thus an empirical theory about how the brain operates and generates stimulus-induced or task-evoked activity thereby presupposing a prediction model of brain. The relevance of predictive coding and its prediction model of brain extends beyond the merely empirical domain of neuroscience, however. It is commonly taken to have significant epistemic implications. The main issue is whether allegiance to the empirical doctrine of predictive coding entails that the brain be construed as an inferentially secluded system.
Jacob Hohwy (2013, 2014) argues that predictive coding entails a self-evidencing brain that has no direct contact with the world and is therefore closed to and inferentially secluded from the world. Others disagree. For instance, Andy Clark (2012, 2013), argues that predictive coding is compatible with a conception of the brain as an open system. This is a tricky issue. One and the same neuronal mechanism, predictive coding, is associated with different and seemingly contradictory characterizations of the brain.
The first aim of this chapter is to argue that the brain can and should be characterized as both an open and closed system. These need not be taken to be contradictory descriptions if sufficient attention is paid to the differences between spontaneous (or resting state) and stimulus-induced activity in the brain. I use the concepts of spontaneous activity and resting state more or less interchangeably (see chapters 1 and 2 for more details on this point). The term resting state is usually used as an operational term denoting a behavioral condition, that is, the state of the brain in the absence of any specific tasks or stimuli. The term spontaneous activity emphasizes that such resting-state activity is not just independent of external stimuli but also generated by the brain itself.
Main Aims and Arguments
It is commonplace to think of the brain as subject to stimulus-induced activity—neural responses to events in the external world. That the brain is also subject to neural activity that is not related to specific stimuli or tasks but rather originates spontaneously within the brain itself, is less well known and too often neglected in discussions about the philosophical implications of neuroscience. I argue that this latter form of neural activity, so-called resting state, or spontaneous activity, is a means by which the brain references its own activity to elements of the external world—this provides the basis for a description of the brain as in part an open and world-evidencing system that includes the world–brain relation and the prediction model of brain.
The second aim of the chapter is to present some empirical facts about an extreme case wherein the resting state’s alignment to the world is altered in an abnormal way. This extreme case is the brain of a schizophrenic, which can serve as an apparent counter-example to the description of the brain as an open system. The main question is whether some symptoms of schizophrenia, such as delusions and hallucinations, should be interpreted as complicating the picture of the brain as an open system with world–brain relation.
In the first part of the chapter, I show that the brain’s resting state aligns to the world in a statistical way. Although I take this to be sufficient permission for deeming the brain an open system, it does entail that different brains can display different degrees of alignment to the world, that is, different types of world–brain relation.
In extreme cases such as schizophrenia, the resting state’s statistically based alignment to the world can break down dramatically. The resulting hallucinations and delusions can be usefully investigated as a clinical analogue to the systematic deception that Descartes imagined perpetrated by an “evil demon” and that appears in the debate between Hohwy and Clark. I conclude that because schizophrenia is an abnormal condition involving abnormal brains, there is good reason not to generalize such epistemic worries across the class of brains in general. In contrast, I suggest that the occurrence of schizophrenic hallucinations and delusions be interpreted as indicative of the fact that the brain’s openness to the world is a highly intricate and therefore delicate phenomenon.
Part I: Predictive Coding and Stimulus-Induced Activity
Predictive Coding Ia: Actual and Predicted Inputs and the Prediction Error
Many early neuroscientists tended to conceive of the brain as mainly in the business of responding to stimulations from the outside world. This resulted in an emphasis on stimulus-induced or task-evoked activity (see also Raichle, 2009, for an overview). One of the things that functional imaging techniques such as electroencephalography (EEG), magnetoencephalography (MEG), positron emission tomography scan (PET), or functional magnetic resonance imaging (fMRI) can do even in the absence of sophisticated theories about neural activity is indicate how the brain responds to stimuli. Thus, many imaging studies seek to show how the brain’s activity changes in response to particular conditions, such as the presence of a visual stimulus in the form of a picture. It was commonly assumed that such stimulus-induced activity was determined mainly, if not entirely by the features of the relevant stimulus. Raichle (2009, 2015a,b) described this framework as an “extrinsic view of the brain” (see also Northoff, 2014a).
This traditional view of the brain’s stimulus-induced or task-evoked activity has been placed into doubt by predictive coding. Specific stimuli or tasks are no longer considered sufficient by themselves to account for the activity changes associated with the brain’s response to stimuli. Instead, what we observe as activity change during stimulus-induced or task-evoked activity results from a process whereby the brain generates predictions of impending input and compares this content to the actual input it receives from the world.
The predicted input is called the empirical prior. Once the actual stimulus arrives, it is set and compared against the empirical prior; if the actual stimulus is identical to the predicted one, the former will not induce any activity change; if in contrast, the actual stimulus diverges from the predicted one, the former will induce strong activity change. The resulting activity change, that is, stimulus-induced or task-evoked activity, thus reflects prediction error, the degree to which the actual input deviates from the predicted input.
Some of the clearest examples of predictive coding occur in the visual cortex (see, e.g., Alink, Schwiedrzik, Kohler, Singer, & Muckli, 2010; Egner, Monti, & Summerfield, 2010; Langner et al., 2011; Rauss, Schwartz, & Pourtois, 2011; Spratling, 2010, 2012a,b). I discuss here one representative study by Rao and Ballard (1999). They approached the question of whether higher visual cortical regions carry predictions for lower ones in terms of feedback connections. The basic idea is that when neural activity in a lower visual area is dependent on that of a higher region, one can safely assume that the latter carries a predicted input for the former. In order to test this assumption, Rao and Ballard applied a computational simulation model of neural activity in lower and higher visual regions.
This allowed them to test the mathematical description of predictive coding. At the time, it would not have been possible to carry out a comparable investigation using functional brain imaging and nonsimulated data. Rao and Ballard (1999) (see also Doya et al., 2011, for an interesting extension concerning decision making) demonstrated that feedback connections from higher to lower cortical areas carry predictions of visual input, which is processed by lower regions’ activities. Lower regions refer to those regions that are more proximally responsive to stimuli whereas higher regions refer to parts of the visual cortex that are more distally responsive.
The lower regions include the primary visual cortex, or V1, where information is received from subcortical regions, such as the lateral geniculate nucleus. This process is followed by subsequent processing of the same visual stimulus in the secondary visual cortex, or V2, a “higher” region. These higher visual regions seem to carry signals that aim at anticipating the incoming visual stimuli processed in the lower regions, for example, V1. The “anticipations” generated in the higher regions are a part of feedforward connections, which are involved in processing discrepancy between the predictions and the actual sensory input.
These findings from the visual cortex show that the brain generates predicted input that is then compared with actual input. This should clarify the sense in which the doctrine of predictive coding takes actual stimuli from the environment as insufficient to account for stimulus-induced or task-evoked activity. The actual stimulus is a necessary but nonsufficient condition that has to be complemented by a predicted input (or stimulus or task) to generate stimulus-induced or task-evoked activity.
Predictive Coding Ib: Predicted Input and Rest–Stimulus Interaction
The study described above just investigated the visual cortex. Karl Friston (2010) proposes a more general hierarchical architecture that posits bottom-up processing of the actual sensory input in the sensory cortex and top-down processing of the same by more cognitive regions in the prefrontal cortex. In addition to the lowermost and uppermost regions of the sensory and the prefrontal cortex, there are many other regions sandwiched in between, whose interrelations need to be explained.
Friston argues that particular regions can serve as both lower and higher nodes in the hierarchy relative to other levels. One region may serve as a higher level of processing relative to another by contributing predictions concerning activation patterns in the latter. The same region may serve as a lower processing level to another region that generates predictions concerning the former’s operations. Since the same region’s neural activity serves as both predicted input (for the next-lower one) and actual input (for the next-higher one), continuous matching and comparison processes occur between lower and higher regions’ neural activities (Friston, 2010).
These continuous matching and comparison processes occur throughout the whole brain enabling the generation of prediction errors at each processing level (Friston, 2010). Clearly, this makes predictive coding a hugely complicated process. At least one thing is clear, however. The anticipated inputs must be generated prior to processing of the actual input or they would not be in place to function as predictions. This means that the level of activity prior to the onset of a stimulus, the prestimulus resting-state activity, must encode the predicted input. The interaction between predicted and actual input may consequently be described as interaction between prestimulus resting state and the actual stimulus. This has been described as rest–stimulus interaction (Northoff, 2014a; Northoff, Qin, & Nakao, 2010). Although the exact neuronal mechanisms underlying such rest–stimulus interaction are currently far from understood (Huang, Zhang, Longtin et al., 2017; Northoff, 2014a, for first steps), that there is some such interaction and that it is important for predictive coding are well-supported claims.
Predictive Coding IIa: Rest–Stimulus Interaction
Various spatial and temporal metrics can be used to measure resting-state activity. Spatial measures such as fMRI target different neural networks, allowing for measurements based of functional connectivity within the networks themselves as well as between different networks (Cabral, Kringelbach, & Deco, 2013; Menon, 2011; Raichle et al., 2001). Temporally, resting-state activity can be measured in electrophysiological or magnetic activity as with EEG or MEG. These techniques target neural activity changes in different frequency ranges, as well as the phenomenon of cross-frequency coupling, which refers to cases of activity in one frequency range being causally related to activity in another frequency range (Cabral et al., 2013; Engel, Gerloff, Hilgetag, & Nolte, 2013; Ganzetti & Mantini, 2013).
One can also measure the brain’s resting-state activity in psychological terms. The brain’s resting-state activity (especially in the default-mode network, or DMN) has been shown to be associated with mind wandering (Mason et al., 2007), random thoughts (Doucet et al., 2012), or self-generated thoughts (Smallwood & Schooler, 2015). Psychologically, resting-state activity seems to specialize in internally generated mental activity (such as thoughts or imagery) as distinguished from externally generated mental contents (such as perceptions).
A recent fMRI study focusing on the auditory cortex (Sadaghiani, Hesselmann, & Kleinschmidt, 2009) shows that resting-state activity impacts stimulus-induced activity and associated perception of objects and events in the world. The investigators had subjects perform an auditory detection task and presented broadband noise stimuli in unpredictable intervals of 20–40 ms. The subjects had to press a button when, and only when, they thought they heard the target sound; otherwise, they were not to hit the button. This allowed the researchers to compare the neural activity preceding hits with the neural activity preceding instances of subjects failing to detect the target sound.
Interestingly, successful detection was preceded by significantly higher prestimulus activity in the auditory cortex in comparison to misses. This was complemented by another analysis of the same data (Sadaghiani, Poline, Kleinschmidt, & D’Esposito, 2015) wherein it was indicated that certain neural networks such as the DMN showed enhanced functional connectivity prior to the onset of those auditory stimuli that were detected.
Taken together, these studies and others (see Northoff, 2014a; Northoff et al., 2010) show that the resting state exerts a strong impact on the contents of our perception. The resting state’s prestimulus activity level seems to be central in determining the contents that we subsequently perceive (see Hohwy, 2013, 2014, for more details).
Sadaghiani, Hesselmann, et al. (2010) explain that their findings concerning prestimulus activity are compatible with predictive coding. The higher the levels of prestimulus activity, the more likely that a specific predicted input (as distinguished from others) is generated. In contrast, lower levels of prestimulus activity may then be assumed to reflect the generation of ambiguous or vague predicted inputs.
The resulting stimulus-induced activity may then be traced to the interaction between the predicted input, as reflected in the prestimulus activity levels, and the actual input, the auditory tone. The better subjects predicted the auditory tone, the higher their levels of prestimulus resting-state activity, and the more likely they were to detect the tone. This makes it clear that stimulus-induced activity (and its associated behavioral and phenomenal effects) is dependent on the level of prestimulus resting-state activity, which provides one example of rest–stimulus interaction.
Predictive Coding IIb: Predicted Input—Brain as a Self-Evidencing and Closed System
One aspect of the traditional view of stimulus-induced activity mentioned above is the idea that the stronger the actual stimulus, the stronger the degree or amplitude of the resultant stimulus-induced activity. According to this traditional picture, since stimulus-induced activity is exclusively determined by the stimulus itself, the brain and its stimulus-induced activity can be regarded as an open system. Being an open system involves the brain setting or referencing its stimulus-induced activity against the actual stimulus and thus, more generally, the environment or the world. The scenario changes though once one accepts predictive coding.
In the case of predictive coding, stimulus-induced activity is no longer exclusively determined by the actual stimulus but also by the predicted input. Predictive coding entails that stimulus-induced activity depends on the degree to which actual and predicted input match or converge: the more predicted and actual input differ from each other, the stronger the resulting stimulus-induced activity.
If, in contrast, they do not diverge from each other, the degree of stimulus-induced activity will be rather low irrespective of the degree or intensity of the actual input itself. The stimulus-induced activity is consequently no longer set or referenced against the actual input and the environment or world but, rather, against the brain’s activity itself. The brain may consequently be characterized as a closed system.
There is therefore something right about Hohwy’s (2013, 2014) claim that the brain is a self-evidencing system. The degree to which the brain reacts to stimuli related to events and objects in the world is mitigated by the brain itself. The brain’s ongoing resting-state activity is as important a contributor as the objects or events of perception themselves.
Thus, if the case made above for the claim that resting-state activity constitutes the predicted input is accepted, then there is reason to believe that the brain is indeed a self-evidencing system that is operationally closed to and inferentially secluded from the world. If we were to stop here, it would seem fair to say that the world’s objects and events can at best impact the brain in an indirect way.
This characterization of the brain as a closed and self-evidencing system has major epistemic implications. The fact that the brain’s predictive coding is based on closed and inferentially secluded processes opens the door for skepticism. It remains impossible for us to rule out the possibility that the objects or events we perceive are more related to the predicted input and hence to the brain’s resting state (and its spatiotemporal structure) rather than directly to the objects or events themselves. Such an internalist model of knowledge may be most visible in extreme cases where the predicted input completely overrides the impact of the actual input.
The overriding of the actual input by the predicted input seems to occur for instance in patients with schizophrenia who suffer from delusions and hallucinations. The events or objects in both delusions and hallucinations are internally generated in the brain’s resting state in the form of predicted inputs that seem to be abnormally strong such that they override (and ultimately preempt) the would-be impact of the actual input (Adams, Stephan, Brown, Frith, & Friston, 2013; Corlett, Taylor, Wang, Fletcher, & Krystal, 2010; Corlett, Honey, Krystal, & Fletcher, 2011; Fletcher & Frith, 2009; Fogelson, Litvak, Peled, Fernandez-del-Olmo, & Friston, 2014; Ford et al., 2014; Horga, Schatz, Abi-Dargham, & Peterson, 2014; Jardri & Denève, 2013; Notredame, Pins, Deneve, & Jardri, 2014).
For instance, in the case of auditory hallucination there is solid empirical evidence for increased resting-state activity in the auditory cortex (Northoff, 2014b; Northoff & Qin, 2011). Such increased resting-state activity may lead to the constitution of abnormally strong predicted inputs that are no longer impacted by any actual input anymore. The internally generated predicted inputs may be so strong as to operate as quasi-actual input with the result that subjects hear illusionary voices, that is, experience auditory hallucinations. The brains of schizophrenics, thus, confuse predicted and actual input, taking the former for the latter. These patients no longer react much to externally occurring inputs, paying low amounts of attention to external sounds while preserving a keen focus on the hallucinated voices.
This theory is supported empirically by the observation that stimulus-induced activity related to external auditory stimuli is abnormally low in some schizophrenics (Northoff, 2014b; Northoff & Qin, 2011). In this situation, the resting state itself constitutes a rather strong predicted input that cannot be modulated anymore by external auditory input. The resulting prediction error, that is, the “stimulus-induced activity” in the auditory cortex, consequently reflects mainly the predicted input with just a marginal contribution from external auditory stimuli. The subjects therefore perceive contents encoded by predicted input, rather than from the external auditory input.
What does such aberrant predictive coding in schizophrenia imply for characterizing the brain? The brain in schizophrenia may indeed be closed to the world (in their world–brain relation) to a higher degree than in healthy subjects. This involves a breakdown of the normally functioning indirect inference of events and objects carried out through predictive coding. For schizophrenic patients, the balance between internally generated predicted inputs and externally generated actual input is shifted abnormally toward the former.
What in the normal case serves as an internally generated reference, for example, the predicted input, against which the externally generated events and objects, including the actual input, is matched and compared, operates now as actual input by itself. The brain’s neural activity is consequently closed to the world to a higher degree than in healthy subjects where the predicted input can still be modulated and impacted by the actual input.
Part II: Prediction Model and the Spontaneous Activity’s Statistically Based Alignment to the World
Prediction Model Ia: Spontaneous Activity and the World–Brain Relation
I have so far traced the predicted input back to the resting state and its specific spatial (e.g., relations between networks) and temporal (e.g., relations between low and high frequencies) features. But how are the resting state itself and its spatiotemporal structure generated and shaped? It is important to address this question since the predicted input is generated by the resting state. The resting state’s spatiotemporal structure including its origin should consequently surface in the predicted input itself.
Although the resting state’s various spatial and temporal features are present in the adult brain, they are only present as predispositions in the infant brain. This means that the resting state’s spatiotemporal structure is strongly experience-dependent one (see also Duncan et al., 2015; Nakao, Bai, Nashiwa, & Northoff, 2013; Sadaghiani & Kleinschmidt, 2013). The concept of experience-dependence means that features of the resting state and its spatiotemporal structure are shaped by the experiences of subjects. For instance, early developmental experiences may have a major impact on the spatiotemporal structure of the resting state.
A recent study showed the resting state’s spatiotemporal structure in adulthood to be predictive of subjects having incurred childhood trauma (Duncan et al., 2015; Nakao et al., 2013). Specifically, the degree of entropy (i.e., the degree of disorder in neural activity across time) in the resting state of adults predicted the degree of early childhood trauma: the higher the degree of early childhood trauma, the higher degree of entropy in the resting state’s spatiotemporal structure in adulthood (Duncan et al., 2015).
This shows that early experiences can be encoded into the resting state’s spatiotemporal structure and can persist thereafter for rather long time frames. The resting state and its spatiotemporal structure may consequently be likened to a mirror of our experience with the world and may therefore be characterized as “experience-dependent.” Such “experience-dependence” of the brain’s spontaneous activity is possible only when it is continuously linked or coupled and thus related to the world. This phenomenon constitutes what I describe as the “world–brain relation.”
Prediction Model Ib: The Spontaneous Activity’s Statistically Based Alignment to the World
In order to encode life events into its spatiotemporal structure, the brain’s resting-state activity must somehow align to these events. What are the neural mechanisms of such alignment? To discuss this, I will focus on a study by Stefanics et al. (2010). These authors conducted an EEG study in healthy human subjects. Subjects were presented with target tones to which they had to react by pressing a button, thus yielding a reaction time.
Preceding the target tone, the investigators presented different cue stimuli (also tones, although with a different frequency than the target tone) that indicated the probability of the subsequent target tone’s occurrence. In the first experiment four different cue tones were presented, one indicating 10 percent; the second, 37 percent; the third, 64 percent; and the fourth a 91 percent probability of the target tone’s occurrence. Depending on the degree of probability indicated by the cue tone, it was followed either by a target tone or by another cue tone a certain percentage of the time.
Following previous data from Schroeder and Lakatos (2009a,b), the authors focused on slow-frequency oscillations in the delta range and their entrainment of faster-frequency oscillations (such as gamma). This approach was adopted because the investigators suspected the slow-fast-frequency entrainment to be related to the statistical probability of the stimulus’s occurrence across time.
What were the findings of Stefanics and colleagues (2010)? As expected, they demonstrated that the reaction time (time needed for the response to target tones) was significantly faster in those trials (target tones) where the preceding cue tones correlated with higher probability. The higher the probability indicated by the cue tone, the faster subjects were able to react. This pattern was observed in both experiments. The subjects thus had been able to learn the probability of the tones.
Does this entail predictive coding with the generation of a predicted input? In order to demonstrate that, one would need to omit the tone at some instances. If the subjects then still showed the same behavioral and neural reaction as in the presence of the tone, those reactions must then be based on a predicted input (since there would be no actual input). The data do indeed confirm that assumption as is discussed below. First, I want to briefly discuss the EEG data.
The EEG data show that the phase of delta oscillation was significantly shifted and aligned, or entrained as is said in neuroscience, to the onset of the target tone as manifest in a significant phase preference. The target tone’s onset was especially locked to the negative phase, that is, the negative deflection in the ongoing cycle of the fluctuations in the delta range. The phase locking was much higher in response to the cue tones indicating higher probability of subsequent target tones.
How is such phase locking possible? It is possibly only if the phases of the delta oscillations actively shift their onsets toward the predicted or expected onsets of the target tone. This is indeed confirmed by the data that showed that higher predictability of the target tone’s onset as indicated by the cue tone induced higher degrees of phase shifting of the delta oscillation’s phase onset. Such a relation between the phase shifting and the predictability of the target tone suggests that the phase onsets are aligned, that is, entrained by the probability of the target tone rather than by its actual presence. The higher the probability of the target tone, the more likely the phase shift occurs irrespective of whether the target tone turns out to occur or not.
However, how is it possible that the phase shift is dependent on the probability rather than the presence of the tone? It is possible because the phase shift reflects the prediction of the actual input, the predicted input, rather than presence of the actual input itself. By indicating higher probability of the target tone, the cue tone makes it easier for the brain to generate a proper prediction, the predicted input, which neurally is manifested in the observed phase shift (see van Atteveldt, Murray, Thut, & Schroeder, 2014).
More generally, one may say that the phase onset of the delta oscillations followed the expected natural statistics of the target tone. Different probabilities of the target tone’s occurrence led consequently to different degrees of phase shifting. These results thus provide empirical support for the claim that the resting state encodes the probability of stimuli in the world and as such accounts for the world–brain relation.
By shifting the phase onsets of especially low-frequency fluctuations such as delta oscillations, the resting-state activity can encode the statistically based temporal (and spatial) differences between different stimuli (see van Atteveldt et al., 2014, for an overview of other pertinent results). If that is true, one would expect that a high probability cue tone without the subsequent presence of the actual tone should lead to the same behavioral and neural reaction as occurs on those occasions when the target tone does occur. This is indeed the case as demonstrated by a second experiment in the study by Stefanics et al. (2010), which is discussed below.
Prediction Model IIa: Brain’s Resting State as Open and World-Evidencing System—Stochastically Based World–Brain Relation
Empirically, one may consider the delta phase shifting to be an example of stimulus-induced activity rather than resting-state activity. Each auditory tone induces stimulus-induced activity that can be traced to rest–stimulus interaction between predicted and actual input and the resulting prediction error (see van Atteveldt et al., 2014, for such an interpretation). If so, delta phase shifting would not add anything new to the interpretation of the brain’s predictive coding activity as exemplary of a closed system. However, this sort of phase shifting is related to the resting state rather than stimulus-induced activity.
The results described above were demonstrated in the first experiment by Stefanics et al. (2010), where the prediction and thus the expected stimulus onset fell together with the onset of the presentation of the target tone. Hence, it remains impossible to disentangle the effects of the resting state from those induced by the target tone itself. To address this, the investigators conducted a second experiment.
The second experiment presented the same target tone but now varied its temporal relation to the cue tones by presenting the target tone either early, right after the cue tone, or rather late. Both early and late target presentations were preceded by two different cue tones that either indicated 20 percent or 80 percent target-tone occurrence. This allowed the experimenters to investigate especially the late-target tone trials when an early target tone was expected (with especially high probability of 80 percent) but not delivered.
In those trials where a cue tone indicating high probability (80 percent) was followed by a late target tone, delta oscillations were locked in their phase to the expected onset of the target even though it was not delivered (because it was a late–target tone trial). Such delta-phase entrainment was observed in conditions where cue tones (20 percent, 80 percent) were followed by late target trials (rather than early target trials). And as in experiment 1, the phase locking to the expected target tone onset was significantly higher in those trials with high-probability cue tones (80 percent) when compared to those with low-probability cues (20 percent). The delta oscillations’ phase onsets were thus shifted to the expected target-tone onsets, even if they were not actually delivered.
What does this experiment tell us about the predicted input? By being exposed to the prior cue tones, the resting state generates a predicted input which, if sufficiently strong in the degree to which the input is predicted, will be exerted even in the absence of an actual tone. Neuronally, this is realized by the delta phase shift in the in those trials where the target tone is actually not presented.
Most importantly, this and other studies (see Northoff, 2014b, as well as van Atteveldt et al., 2014, for details) support the view that the resting state’s spatiotemporal structure (as for instance its delta phase onsets) are statistically based (rather than one single stimulus). The resting state’s delta phase in particular and its temporal structure in general are based on the statistical occurrence of myriad stimuli across time in the environment rather than based on the occurrence of a single stimulus at one particular point in time. This amounts to a stochastically based world–brain relation.
Prediction Model IIb: Brain’s Resting State as Open and World-Evidencing System—Empirical and Conceptual Confusion?
The data about delta phase shifting can only be obtained by considering several stimuli and, more specifically, their statistical frequency (probability) distribution across time. A longer time scale such as this differs significantly from those presupposed in the data by Stefanics et al. (2010) and the other data about prestimulus effects described earlier. There, the time scale is extremely short covering only single stimuli while neglecting the impact of different stimuli in the brain’s neuronal activity across time.
Quite generally, studies of stimulus-induced activity focus on the neural activity related to single stimuli at a particular point in time thus neglecting resting-state activity, the study of which requires observing the brain’s response to several stimuli over time. Due to its dependence on the statistically based occurrence of stimuli across time, phase shifting in general and delta-phase shifting in particular should be associated with resting-state activity (see also Klein, 2014).
Single stimuli are usually presented for several milliseconds or seconds that impact neural activity within this rather short time frame. Resting-state activity, in contrast, presupposes a much longer time scale. As we have seen in the previous section, resting state is not restricted to one particular brief stimulus mimicking a single auditory tone at one particular point in time. Instead, resting-state activity seems to encode information concerning a much longer time scale that extends over several tones (as in the experiment by Stefanics et al., 2010) or even years (as in the childhood trauma data). This makes it clear that neural activity in resting state and stimulus-induced activity operate on different time scales.
What does this imply for our characterization of the brain in general? The resting state aligns itself in terms of an “extra-neuronal loop” (Clark, 2013) to the statistically based temporal and spatial structure of objects and events in the world. It is not, therefore, closed and inferentially secluded when referring to and measuring its own resting-state activity against input from actual stimuli. Andy Clark, one of the main proponents of the view of the brain as an open system, also emphasizes the statistical nature of the brain’s neural activity and bases his account of predictive coding on it (Clark, 2013, pp.4 and 8). He does not yet make the distinction, however, between resting state and stimulus-induced activity, which puts his account of predictive coding in direct opposition and seemingly logical contradiction to that of Hohwy (2013, 2014) (see figure 3.1).
It may be odd, but it is not contradictory for me to have characterized the brain as both open and closed to the world. This is not contradictory once one embraces the distinction between resting state and stimulus-induced activity. The brain’s stimulus-induced or task-evoked activity as its response to specific stimuli or tasks is indeed closed to and inferentially secluded from the world: it is generated in relation to the brain’s predicted input and its prestimulus resting-state activity level reflecting an “intraneuronal loop.”
In contrast, this does not apply to the brain’s resting-state activity that is aligned to the world in a statistically and spatiotemporally based way. The brain’s resting-state activity is consequently aligned and open to the world because it is part of an “extraneuronal loop” as Clark (2013) describes it. By looping outside itself into the world, the brain (and its resting state as I would add to Clark’s case) becomes open to the world.
Prediction Model IIIa: The Brain’s Openness to the World—Empirical and Epistemological Meanings of “Open”
Why is the encoding of the statistical occurrence of stimuli by resting-state activity important in this context of interpreting predictive coding? The predicted input is generated by the resting state and, more specifically, the prestimulus resting state. That by itself can lead to the adoption of the view that the brain is a closed and self-evidencing system. However, if the resting state itself encodes the statistical occurrence of stimuli, perhaps it should not be conceived of as a closed system. Instead, the resting state may then be considered an open system that references or sets its own activity level against the statistical occurrence or probability of stimuli in the world or in the environment.
However, we have to be careful here with what exactly we mean by “open.” The concept of “open” can be taken in an empirical or epistemological sense. If taken in a purely empirical way, it simply means that the brain’s resting-state activity by itself has direct access to the external world in that it encodes the statistical frequency distributions of external stimuli from the world. The question guiding the empirical meaning of “open” is: “How is the brain’s spontaneous or resting-state activity related to external stimuli in the world?”
This purely empirical meaning of “open” is much weaker than its epistemological sibling. The epistemological meaning of “open” entails not only direct contact (as in the empirical meaning) but, much stronger, that what is encoded into the brain’s resting-state activity reflects the truth. The question guiding the epistemological sense of “open” is: To what extent should a process of statistical alignment between the predicted inputs of the brain and events in the world be taken to confirm the world–brain relation as truth preserving?
The distinction between empirical and epistemological meanings of “open” is especially important in the philosophical context. The philosopher concerned by the Cartesian demon might reject the openness of the mind, in the epistemological sense of “open.” The empirical and epistemological meanings of “open” can dissociate from each other, however. The case of schizophrenia already described is an instance where a degree of empirical openness is accompanied by seemingly more pressing epistemological closedness. The brain, despite being empirically open, can fail to be world-evidencing in severe schizophrenia. Although empirically open, the schizophrenic’s brain may nevertheless be closed epistemologically. This complication is explored below.
Prediction Model IIIb: The Brain’s Openness to the World in Schizophrenia—Exclusion of the Cartesian Demon?
I pointed out that delta phase locking is central in aligning the brain’s spontaneous activity to external stimuli in the world. However, what if the described phase locking and thus the resting state’s “extra-neuronal loop” do not function properly anymore? That seems to be the case in schizophrenia (Lakatos, Schroeder, Leitman, & Javitt, 2013). Using essentially the same design as Stefanics et al. (2010), Lakatos et al. (2013) recently conducted an EEG study in schizophrenic patients to whom they presented a stream of auditory stimuli (i.e., tones) with regular, that is, rhythmic interstimulus intervals (1500 ms). The stream of auditory stimuli included some deviant stimuli (20 percent) that were distinguished in their frequency. Healthy and schizophrenic subjects had either to listen passively (passive task), detect the easily detectable deviant stimuli (easy task), or detect the more difficult (variation by frequency) detectable stimuli (difficult task).
Schizophrenic patients did show a much lower degree of delta phase locking in response to the onsets in the stream of auditory stimuli when compared to healthy subjects. Moreover, the degree to which delta phase locking was impaired correlated with the severity of a subject’s psychopathological symptoms of hallucinations and delusions. This strongly suggests that the degree of phase alignment or entrainment is closely related to the kind of objects or events one perceives. If the resting state can properly align itself to the statistically based temporal (and spatial) structure of objects and events in the world, the latter form the content in one’s perception. In that case the looming threats of the actual occurrence of a Cartesian demon and skepticism are more or less (that is, in a statistically based way) excluded.
If, in contrast, the resting state can for some yet unclear reason no longer properly align itself to the objects and events in the world, the contents in perception and cognition become detached and dissociated from the events and objects in the world resulting in delusions and hallucinations as in schizophrenia. In that case the danger of a Cartesian demon resurfaces. This means that due to the statistically and spatiotemporally based nature of the resting state’s alignment to the world, we remain unable to principally exclude the possibility of a Cartesian demon-like drift away from the world’s objects and events. However, this only applies to extreme cases as in schizophrenia. In the “normal and healthy” case, our brain’s resting state can align us to the world in a statistically and spatiotemporally based way: this amounts to a stochastically based world–brain relation.
Importantly, this means that the characterization of the brain’s resting state in particular and the brain in general by openness to the world due to an “extraneuronal loop” does not necessarily exclude the remote possibility of a Cartesian demon-based skepticism. Still, we may consider the skepticism issue in a statistically based way. Statistically our brain’s spontaneous activity or resting state aligns us to the world and keeps us open to its events and objects in most instances. Thus, taken in a statistical sense, our cognition and knowledge of the world can be secured in the majority of instances. This may be relevant for future epistemological discussion that then may also determine the concept of skepticism in a more detailed and sophisticated way.
The possibility of both the Cartesian demon and skepticism (in whatever epistemological–philosophical version) consequently rests ultimately on the statistically based degree to which our brain’s resting-state activity is aligned with and open to the world. If the resting state’s degree of statistically based alignment to the world tends toward zero as in the case of schizophrenia (with, for instance, low degrees or absent delta-phase shifting), the predicted inputs generated by the resting state no longer carry predictive value but are, instead, neurally mistaken for actual inputs. If, in contrast, our brain’s spontaneous activity or resting state aligns us well to the world (with for instance high degrees of delta-phase shifting), our predicted inputs generated by our resting state carry a high degree of fidelity that minimizes the possible basis for Cartesian skepticism.
How does this affect our knowledge? Our knowledge of the world depends on the brain and its statistically based relation to the world, for example, the world–brain relation. On the basis of such a world–brain relation, we can generate models about the world including the world–brain relation. Those very same models (see chapters 2 as well as 9 and 10) are based on both brain, that is, brain-model dependent, and world, that is, world–model dependent (see chapters 9 and 10).
If our knowledge were only and exclusively dependent on the brain itself, for example, brain-model dependent, independent of world–brain relation and concurrent world-model dependence, the door for skepticism would be wide open. This is not the case however. Our brain’s spontaneous or resting-state activity aligns itself to the world’s statistical frequency distributions, or, the world–brain relation, which entails world-model dependence (see chapter 10 for details). That closes the door for skepticism and keeps it only minimally open, that is, in a statistical sense with low probability. This hinges strongly on the relationship between ontological and epistemological assumptions, an issue that will be central for reformulating the mind–body problem as the world–brain problem (see chapters 12–15).
Conclusion
I introduced predictive coding with predicted input and prediction error as one of the major models of the brain’s neural activity in current neuroscience. On the level of model, this entails a prediction model of brain. The prediction model of brain raises the question whether the brain’s neural activity is closed with respect to itself and thus self-evidencing, or, alternatively, whether the brain’s neural activity is open to the world and thus world- rather than self-evidencing. The empirical data support the view of the brain as an open and world-evidencing system: these show that the brain’s spontaneous or resting-state activity is aligned in a statistically based way to the statistical frequency distributions of events in the world, the world–brain relation.
Does our brain, with world–brain relation and the prediction model, open the door for the Cartesian demon and skepticism (in whatever version) in our cognition and knowledge of the world? The empirically informed philosopher may say that in the normal or healthy case, our cognition and knowledge reflect more or less, that is, in a statistically based way, the world as it is by itself. This is due to the brain and its spontaneous activity’s or resting-state’s statistically based alignment to the world. That however, so the traditional philosopher might claim, does not exclude the theoretical possibility of skepticism.
Accordingly, to reconcile both the empirically informed and the traditional philosopher, we cannot exclude in principal that our cognition and knowledge are infected by the Cartesian demon and skepticism. This however is exceedingly unlikely in the “normal” case but possibly prevalent in extreme cases such as those of schizophrenia. Accordingly, we cannot exclude the Cartesian demon and skepticism in an absolute sense. However, we can nevertheless say that, on statistical grounds, the probability of such a Cartesian demon and its skepticism is rather low if not minimal, except when one is in an extreme state such as schizophrenia.
