As detailed in several previous reviews the neurosciences are about to undergo a paradigm shift towards concepts that consider the brain as a self-organizing complex system with non-linear dynamics that exploits a huge body of stored knowledge to interpret sensory signals, formulate hypotheses and to generate predictive models of the world in order to optimize adapted behavioral responses (Singer 2009, 2013). For many decades, the search for the neuronal underpinnings of cognitive and executive functions has been guided by the behaviourist view that the brain is essentially a highly complex and versatile stimulus-response machine, in which serial processing strategies prevail. This view received further support from early anatomical data that emphasized that feed-forward connections exhibit high topographic precision and possess strong driving synapses, while feedback connections are diffuse and only modulatory. The rather impressive performance of artificial pattern-recognition systems based on such processing architectures suggested that neuroscientists were on the correct path. Accordingly, they set out to study the responses of neurons to sensory stimulations across the various stages of the processing hierarchy and analyzed activation patterns associated with motor output, hoping that these strategies would eventually lead to reductionist explanations of the neuronal mechanisms that support cognition, memory, decision making, planning, and motor behaviour. The strategy to follow the transformation of activity from the sensory surfaces over the numerous levels of hierarchically-organized processing structures to the respective effector organs proved to be extremely fruitful. Comparison of brains from different species provided compelling evidence that the basic principles according to which neurons function and exchange signals have been preserved throughout evolution with only minor modifications. For the comparatively simple nervous systems of certain invertebrates, this behaviourist approach allowed for near-complete descriptions of the neuronal mechanisms underlying particular behavioural manifestations. This nurtured the expectation that pursuing this research strategy would sooner or later allow us to explain in the same way the more complex behaviour of mammals—and ultimately also the highly-differentiated cognitive functions of primates and human subjects. However, in recent decades the pursuit of this approach has led to an accumulation of evidence that demands a revision of the classical hypothesis, which emphasizes serial feed-forward processing of sensory information within hierarchically-organized architectures.
Advances in the analysis of the cortical connectome, the introduction of multisite recording techniques, and the development of imaging methods assessing whole-brain activity have generated data that necessitate an extension of classical views, raise novel questions, and likely provide new solutions to old problems.
Anatomical evidence: i) Within processing streams from sensory surfaces to executive organs, feedback projections are in general more numerous than feed-forward projections, emphasizing the importance of top-down control (Felleman & van Essen 1991). ii) Connections linking neurons within distinct cortical areas cross the boundaries between areas (Schwarz & Bolz 1991). Thus, the cerebral cortex appears to be a continuously coupled sheet, the different cortical areas being distinguished mainly by their input and output connections. iii) From primary sensory areas onwards, processing streams diverge into numerous parallel pathways whose nodes are linked by massive reciprocal connections, both within and across modalities (Markov & Kennedy 2013; Markov et al. 2013). iv) The rule that feed-forward connections originate in supra- and feedback projections in infragranular layers does not hold for nearby cortical areas (Markov & Kennedy 2013; Markov et al. 2013). Together with electrophysiological evidence (De Pasquale & Sherman 2011), this threatens the strict distinction between feed-forward driving and feedback modulatory connections. v) Finally, statistical analysis of interareal connectivity suggests an organisation resembling small-world, rich-club networks (see Van den Heuvel & Sporns 2013) that minimize path length between nodes (areas; Van den Heuvel & Sporns 2011; Sporns 2013). However, analysis of projections with cellular resolution suggests as one reason for short path length the surprisingly high degree of connectedness among cortical areas. Statistical analysis suggests that more than 60% of possible links between network nodes are actually realized (Markov & Kennedy 2013).
Functional evidence: i) Even in early sensory areas neurons lose their simple feature-specific responses when challenged with complex stimuli (David et al. 2004; Vinje & Gallant 2000). Moreover, responses are influenced by stimuli in other modalities, by attention, reward expectation, and contents in working memory, thus suggesting contextual modulation not only by intrinsic connections but also by top-down projections (Engel et al. 2001; Calvert et al. 1997; Iurilli et al. 2012; Muckli & Petro 2013; Stokes et al. 2013). ii) The notion of strictly serial processing from input layer four, via layers three and two, to the output layers five and six of the cerebral cortex needs to be revised in light of evidence that vigorous responses can also be elicited by sensory input when parts of this canonical circuit are disrupted (Constantinople & Bruno 2013). The possibility that supra and infragranular compartments can operate in parallel is further supported by evidence that the two subdivisions engage in oscillatory activity in different frequency bands (gamma in supra- and alpha or beta in infragranular layers; Buffalo et al. 2011; Roopun et al. 2008. iii) Multisite recordings indicate that “spontaneous” fluctuations in the responsiveness of individual neurons are often the reflection of coordinated, highly structured spatio-temporal activity patterns rather than the result of noise (Kenet et al. 2003; Fries et al. 2001a). iv) Widely distributed cortical areas exhibit coherent fluctuations of their spontaneous activity, forming functionally-coupled networks that change in their composition in a state-dependent way (Fox et al. 2005; Hipp et al. 2012; Raichle 2011; Raichle et al. 2001). Thus, the cortex—and in a wider sense the brain—appears to be a highly active, pattern-generating system, rather than just a stimulus-driven device. v) Analysis of whole brain activity with functional magnetic resonance imaging (fMRI) and electroencephalographic (EEG) and magnetoencephalographic (MEG) measurements indicates that virtually all cognitive and executive functions are associated with the activation of networks of often widely-distributed cortical areas (Engen & Singer 2013; Friederici & Gierhan 2013; Hipp et al. 2011; Hodzic et al. 2009; Power & Petersen 2013). This suggests that distributed networks are a substrate of functions rather than individual specialized structures. vi) Finally, analysis of the brain’s dynamic signatures indicates that neuronal populations can engage in oscillatory activity in characteristic frequency bands and synchronize their discharges, such that the respective frequency bands and the composition of coherently active cell groups depend on central states, attention, cognitive tasks, and goals of action (Buzsáki 2006; Singer 2010).
These novel anatomical and functional data suggest as a prevailing organizational principle distributed processing in densely coupled, recurrent networks with non-linear dynamics, which are capable of supporting high dimensional states. This organization requires a high degree of coordination of distributed processes, suggesting that special mechanisms are implemented to dynamically bind local processes into coherent global states, and to configure functional networks “on the fly” in a context- and goal-dependent way. It has also become clear that the brain is by no means a stimulus-driven system. Rather, it is self-active, permanently generating highly structured, high-dimensional spatio-temporal activity patterns. These patterns are far from being random, and instead seem to reflect the specificities of the functional architecture that is determined by genes, modified by experience throughout post-natal development, and further shaped by learning. These self-generated activity patterns in turn seem to serve as priors with which incoming sensory signals are compared. Perception is now understood as an active, reconstructive process, in which self-generated expectancies are compared with incoming sensory signals. The development of methods that allow simultaneous registration of the activity of large numbers of spatially-distributed neurons revealed a mind-boggling complexity of interaction dynamics—which in turn eludes the capacity of conventional analytical tools and, because of its non-linearity, challenges hypotheses derived from intuition.
In the last decade theoreticians have begun to explore and appreciate the immense computational power of such self-organizing recurrent networks that gave rise to concepts such as “reservoir computing”, “echo-state computing” or “liquid computing” (Buonomano & Maass 2009; Lukoševičius & Jaeger 2009). The evidence that resting-state activity is highly structured, that information is contained in the spatio-temporal relations between the responses of widely distributed neurons, and that stimulus-response functions depend crucially on state variables generated within the brain are in principle compatible with such advanced concepts of information processing in highly non-linear, high-dimensional dynamic systems; but neurobiological approaches taking such considerations into account are still very rare.
Our rather detailed knowledge of the response properties of individual neurons in different brain structures, and of the microcircuits that shape these responses, stands in stark contrast to our ignorance of the complex and highly dynamic processes through which the myriads of spatially-distributed neurons interact in order to produce specific behaviours. Evidence from invasive and non-invasive multi-site recordings indicates that most higher brain functions result from the coordinated interaction of large numbers of neurons, which become associated in a context- and goal-dependent way into ad-hoc formed functional networks. These networks are dynamically configured on the backbone of the anatomical connections (for review see von der Malsburg et al. 2010). Evidence also indicates that these interactions give rise to extremely complex spatio-temporal patterns that are characterized by oscillations in a large number of different frequency bands, which can synchronize, exhibit phase shifts, and even cross frequency coupling (Uhlhaas et al. 2009). In the light of these novel data, the brain—and in particular the neocortex—appears to be a self-active, self-organizing “complex system” which exhibits non-linear dynamics, is capable of utilizing multiple dimensions for coding (space, amplitude, oscillation frequency, and phase), operates in a tightly-controlled range of self-organized criticality (Shew et al. 2009; edge of chaos), and constantly generates highly-structured, high-dimensional activity patterns that are likely to represent stored information. However, how exactly information is encoded in the trajectories of these high-dimensional and non-stationary time series is largely unknown, and is the subject of increasingly intense research. Moreover, with the exception of a few studies in which selective manipulation of the activity of defined neuron groups were shown to affect behaviour in a particular way (Salzman et al. 1992; Houweling & Brecht 2008; Han et al. 2011) most of the available evidence on the relations between neuronal responses and behaviour is still correlative in nature. This makes it difficult to determine whether an observed variable is an epiphenomenon of a hidden underlying process or is causally involved in accomplishing a particular function. Thus, systems neuroscience now faces the tremendous challenge of analyzing the principles of distributed dynamic coding and of obtaining causal evidence for the functional role of specific activation patterns, in order to distinguish between functionally-relevant variables and epiphenomena.
In conclusion, we have to abandon classical notions of the neuronal representation of perceptual objects and, in the same vein, that of motor commands. The consequence is that it becomes once again unclear how the distributed processes that deal with the various properties of a perceptual object—its visual, haptic, acoustic, olfactory and gustatory features—are bound together in order to give rise to a coherent representation or percept. Given this, it may appear more than bold to attempt to identify the neuronal correlates of consciousness—probably the highest and most mysterious of our cognitive functions.
Despite the numerous gaps in our understanding of integrated brain functions, neurobiologists agree on a number of general conclusions on the relation between brain processes and behavioural phenomena. The majority of neurobiologists seem to consent that all cognitive and executive functions that we can observe in human beings, including the highest mental activities and consciousness, are the result, not the cause, of neural interactions. Consequently, it is held that mental phenomena follow or emerge from neural interactions and do not precede them. Furthermore, it is assumed that all neural processes obey the known laws of nature. The reason for this is that the behaviour of organisms of low complexity, such as, for example, molluscs or worms, can be fully explained by registering the activity of their neurons and establishing causal relations between the spatio-temporal patterns of this activity and the respective behaviour. There is, at present, no need to postulate any additional unknown forces, laws, or modes of interaction in order to explain their behaviour. The reason for this is that evolution is a very conservative process. Once an invention has been made that increases fitness it tends to be conserved, unless there is a major change in conditions that makes this invention obsolete or maladapted. Therefore, our nerve cells function in exactly the same way as those of snails. Likewise, the development of structures also follows a very conservative path. Since the first appearance of the cerebral cortex, the six-layered sheet of nerve cells that covers the hemispheres of the brain, no new structures have emerged. There is just more of the same, and this increase in complexity marks the difference between the brain of a human being and that of our nearest neighbours, the great apes. Apparently, this processing substrate and the associated gain of complexity marks the difference difference between species that failed and those that succeeded in promoting cultural evolution—with all its far reaching consequences. In this context, however, one needs to consider that cultural evolution created a socio-cultural environment of ever-increasing complexity that in turn contributes to the epigenetic shaping of brain architectures. Thus, even if the genetically-determined layout of brain architectures has changed little since the beginning of human civilisation, those features that can be modified by epigenetic shaping are likely to have undergone major modifications. This fact has not always been taken into account in the past; but its implications will be discussed below. But this additional twist concerns the epigenetic modifiability of our brains, and not its basic functional principles.