The assumption of sparse coding holding on the regional level of the brain raises the question of how functions and regions are related to each other. Historically, a one-to-one relationship between function and region has often been assumed, amounting to what is called “localizationism.” Alternatively, others have suggested that more than one region or network is recruited during one particular function, and that different functions may recruit the same or at least overlapping regions and network—this has been subsumed under the concept of “holism.” I here hypothesize that localizationism and holism are not contradictory but rather complementary. Holism concerns the process level that can be characterized by difference-based coding, while localizationism refers to the outcome or result of the neural processing in terms of differences as described by sparse coding. Since they concern distinct aspects, processes, and the outcome of those processes, holism and localizationism must be regarded as complementary rather than contradictory.
Key Concepts and Topics Covered
Localization, holism, history of neuroscience, overlap of different functions in the same regions/neural networks, sparse coding, difference-based coding, process and outcome, complementarity
Neurohistorical Remark IA: “Localizationism” in Past and Present Neuroscience
One of the main methodological approaches in neuroscience at the beginning of the twentieth century was the investigation of patients with brain lesions. These patients could reveal how their higher-order cognitive functions like consciousness, memory, attention, learning, and so on, were affected by lesions in particular regions.
This was the way that early neurologist Paul Broca found out about a specific region in the brain being in charge of comprehending language—the Broca region. He observed that patients with a lesion in the left lateral prefrontal cortex showed major deficits in uttering words and language, a so-called aphasia. From his clinical observations Broca inferred that this region must be in charge of producing words, thus localizing language in the Broca area, as it is called these days.
Observation of patients with lesions and their corresponding mental disturbances has since been a major tool of insight into the function of the brain. From the exact localization of the lesion and the corresponding mental disturbances, one may infer which region in the brain mediates the respectively underlying higher-order cognitive function. Many other higher-order cognitive functions, including consciousness and self, are currently investigated in this way in neurological patients who suffer from specifically localized lesions in the brain (see, for instance, Feinberg (p.308) 2009; as well as Feinberg and Keenan 2005). This entails what I describe as a “localization-based approach” to the brain.
The concept of the “localization-based approach” can be defined in two ways. First, it implies the neuropsychological assumption that a particular function can be related to the neural activity in a specific brain region, meaning that the former can be localized precisely in the latter. This is a neuronal (or better, neuropsychological) meaning of the concept of “localization-based approach” that pertains to a hypothesis about how the brain’s regions are related to psychological functions.
In addition to such neuropsychological meaning, the concept of the “localization-based approach” can also refer to an investigator’s particular methodological strategy for approaching the brain. The brain here is approached in terms of regions rather than in terms of, say, processes or codes (see Introduction for such code-based approach to the brain). The methodological approach to the brain in terms of regions is not restricted to the investigation of patients with local brain lesions. It may also extend to the healthy subjects, such as, for instance, those investigated in functional magnetic resonance imaging (fMRI). The use of techniques like fMRI is indeed guided by the search for the localization of particular functions in specific regions of the brain, which it therefore approaches in terms of regions (as distinguished from processes or codes).
Finally, the search for localization of higher- order cognitive functions in patients with brain lesions and functional brain imaging converges with the assumption of modules in cognitive psychology. Cognitive psychology proposed specific functional unities that are in charge of processing and operating such specific cognitive content as attentional content, working memory content, conscious content, self-specific content, and so on. When cognitive psychology entered neuroscience and they were amalgamated into “cognitive neuroscience, ” the concept of modules was combined with the concept of localization in the brain (see van Eijsden et al. 2009 for a nice description).
What were described as “modules” in cognitive psychology could then be easily transferred to the brain and more specifically to particular brain regions and their connections. Hence, the localization-based view of brain function seems to be intimately coupled with the module-based view of psychological functioning. This resulted in the assumption of the localization of specific cognitive modules in particular regions (or networks of regions) in the brain.
This is still the implicit or explicit presupposition in current neuroscience and especially in cognitive neuroscience (see, for instance, Logothetis 2008), which is often extended to the more recent branches of affective and social neuroscience: “I take the modular organization of many brain systems as a well-established fact, and discuss only how far fMRI can go in revealing the neuronal mechanisms of behavior by mapping different systems modules and their dynamic interrelationships” (Logothetis 2008).
Neurohistorical Remark IB: Holism in Past and Present Neuroscience
However, nothing in the science of the brain goes without the opposite suggestion. A strictly localization-based approach was put into doubt early on by another neurologist, Hughlin Jackson, who suggested a more complex and systematic neural organization with multiple interdependencies between different regions. This paved the way for a more holistic view of brain function, one that relates higher-order cognitive functions to the neural operations in the whole brain and its multiple regions.
Interestingly, Sigmund Freud, the founder of psychoanalysis, who initially was a neuroanatomist, also rejected a localization-based approach to the brain. His reason was that more complex psychological disorders like hysteria or depression could not be confined to alterations in specific brain regions. He instead regarded these disorders as more complex systems disorders where the organization of the “psychic apparatus, ” as he called it, is abnormal, which is manifested throughout the whole brain and its different regions. One may therefore consider Freud a forerunner of a more holistic view of brain function (see Northoff 2011 and 2012 for details). (p.309)
Later, neuroscientist Karl Lashley (1943, 1950) observed in his postmortem dissections that the extent of a brain lesion predicts the degree to which higher-order cognitive functions and mental states are disturbed. This let him develop what he called the “Law of Equipotentiality” and the “Law of Mass Action.”
Both laws describe the distribution of neural processing across the whole brain during higher-order cognitive functions like consciousness and memory. Different regions were proposed to contribute equally to the generation of complex functions that therefore must be considered the result of “mass action” in the brain. This means that higher-order cognitive functions like memory and consciousness were assumed to result from the neural processing throughout the whole brain, rather than being localized in particular regions or modules within the brain (see also other authors like Koehler and Goldstein as cited in the Introduction).
Analogous observations were made by Russian neuropsychologist A. R. Lurija (1962, 973).
Based on his lesion patients, he suggested that one region in the brain can be involved in various higher-order cognitive functions. Conversely, he postulated that higher-order cognitive functions are mediated not only by one or two regions but by various regions in the brain. Most important, the same higher-order cognitive function may even recruit different regions in different instances, depending on the respective psychological and neuronal contexts. There is thus what Lurija described as “dynamic localization.”
This led Lurija to formulate his hypothesis of functional systems as the operating systems of the brain that describe the actual constellation of different regions that mediate a particular function:
According to this view a function is, in fact, a functional system (...) directed towards the performance of a particular biological task and consisted of a group of interconnected acts that produce the corresponding biological effect. The most significant feature of a functional system is that, as a rule, it is based on a complex dynamic “constellation” of connections, situated at different levels of the nervous system, that in the performance of the adaptive task, may be changed with the task itself may be unchanged. (Lurija 1962)
How about holism in the neuroscience of our days? The earlier-described metabolic approach to the brain by Shulman (van Eijsden et al. 2009) presupposes a more holistic approach to the brain (see Chapter 6). By considering the global metabolic-energetic supply and distribution to the brain as a whole as central for any subsequent neural activity, a holistic, and thus global, component is introduced.
Such a more-holistic view is also promoted in parts of functional brain imaging that focus much more on neural networks spanning across different regions rather than on single regions. This is especially apparent in the functional brain imaging of the resting-state activity (see Chapter 4 for details). However, as we will see further down, even the characterization of the brain by different networks may still presuppose too localizationism.
Finally, the holistic view of the brain also surfaces in the debate about consciousness. As we will see in Volume II, a global workspace of neural activity and information spread is often considered central in constituting consciousness; since such a global workspace allows for global extension and distribution, it implies the involvement of different regions and networks throughout the whole brain (see Introduction and Chapters 18 and 19 in Volume II for details).
Neurohistorical Remark IC: Problems of Localizationism in Present Neuroscience
What is the standing of such a holistic view of brain function these days? The introduction of functional brain imaging has shifted the pendulum back again toward the localization-based view with the assignment of specific regions or networks to particular functions like attention, working memory, and so on (see van Eijsden et al. 2009 for a nice description).
In addition to the various regions and neural networks supposedly serving specific psychological functions, a network particularly involved in mediating resting-state activity, the (p.310) default-mode network (DMN), has been distinguished in regional and connectional terms. The DMN seems often (though implicitly) to be regarded as the module for the resting state that therefore stands side by side with other networks that function as modules for specific functions such as, for instance, executive functions or salience (see, for instance, Menon 2011).
However, recent imaging studies shed some doubt on the proclaimed localization of specific psychological functions in particular regions or neural networks. The various regions of the DMN, like the anterior and posterior cingulate cortex and the medial prefrontal and parietal cortex, are supposed to serve psychological and mental activity, specifically in the resting state (see especially Chapter 26 for details on that). The same regions are also recruited during a variety of psychological tasks or functions, including contextual association, navigation and spatial processing, episodic memory, decision making, execution errors, self-related processing, mind-reading, emotional processing, and social interaction (see Bar et al. 2007, 2009; Spreng et al. 2009).
This sheds some doubt on the regional or network specificity of the DMN; more specifically, on its specific association with particular psychological functions during either resting-state activity or stimulus-induced activity. Conversely, these observations also argue against region-specific (or network-specific) localization of the various functions themselves, which seem to recruit more or less the same regions and networks.
This situation with the recruitment of the same regions and network by different functions is not peculiar to the DMN. The same pattern can be observed in the case of another neural network that includes the bilateral anterior insula, the dorsal anterior cingulate cortex, and the thalamus as its core regions (these regions are also subsumed under what is described as the “salience network”; see Menon 2011). These regions are active during functions as diverse as interoceptive awareness (Critchley et al. 2004; Wiebking et al. 2010), empathy (Yan et al. 2011), anticipation of emotions (Bermpohl et al. 2006), and aversion (see Hayes and Northoff 2011). The list of regions that are recruited by different functions can easily be extended.
In sum, the observation of the same region and network mediating a variety of different functions sheds some doubt upon the localization-based approach and its attempts to establish a specific one-to-one relationship between regions/networks and functions.
Does this mean that we have to revert to a more holistic view of the brain and its different regions? Based on their data, some neuroscientists—doing either lesion-based studies (Feinberg 2009) or functional imaging using electroencephalography (EEG; John 2006), positron emission tomography (PET; van Eijnsden et al. 2009), or functional magnetic resonance imaging (fMRI; Northoff 2008)—do indeed advocate a more holistic view of brain function. This is further corroborated by neuroanatomy, which considers single regions as hubs or nodes within the neural network of the whole brain rather than as centers or modules by themselves (see Hagmann et al. 2008, Sporns 2011).
Where does this leave us? Do we have to follow the swings between localizationism and holism? My aim in the following discussion is to show how both are very compatible and complementary, rather than being contradictory.
Neurotheoretical Remark IA: Localization and Sparse Coding
While the association of a specific region or network with a specific psychological function must be considered doubtful, the data nevertheless show that only a certain set of regions is recruited during the various tasks or functions. Multiple functions seem to recruit the same set of regions or network entailing a many/multiple-to-one/few relationship between functions and regions. The function–region relationship thus seems to obey the rules of sparseness, with sparse representation of the multiple functions in a few regions/networks of the brain. I consequently hypothesize sparse coding rather than localization to operate and determine the function–region relationship.
The assumption of sparse coding is empirically supported by the data we discussed in (p.311) Chapter 3 and especially Chapter 6, which show the spatiotemporal activity pattern during resting-state and stimulus-induced activity to be rather sparse. While these purely neuronal data did not directly address the more neuropsychological relationship between region and function, they nevertheless provide some indirect support for a sparse encoding of functions into the different regions’ and networks’ neural activities.
More specifically, I propose that what is considered localization of a particular function in a specific region reflects the sparse number of actually activated regions when compared to the total number of regions that could possibly be recruited. The fact that the other regions are not activated does not mean, however, that they do not participate in generating the function in question.
The inactive regions may nevertheless have an important role in that their baseline—that is, resting-state activity—may serve to generate and amplify neural differences (presupposing difference-based coding on a regional level; see Chapter 3). These neural differences may in turn allow the brain to condense and sparsen neural activity in one or a few subsequent regions, yielding those regions that we observe to be activated. Accordingly, sparse coding on a regional level seems to be nicely compatible with the localization of particular functions in specific regions.
How does the assumption of such sparse coding stand in relation to the localization approach? To equate sparse coding with localization is to confuse the underlying processes and their resulting outcomes. The localization-based approach focuses on the outcome while neglecting the process itself; that is, how the apparent localization of a function in a particular region is generated. Instead of considering the process of generating regional localization, the localization-based approach takes the localization of a particular function in a specific region for granted. And it considers the psychological function to be intrinsic or innate to the region itself without further questioning the underlying processes how that function is generated by the region’s neural activity.
Such a localization-based approach is, however, to be distinguished from the approach sparse coding takes to the question of localization. Here the focus shifts from the outcome, the observation of a regional localization, to the processes; that is, the rules and principles that generate what we observe as the specific linkage between function and region.
Neurotheoretical Remark IB: Distinction Between “Activated” and “Active” Regions
I discussed the processes underlying sparse coding on a regional level of neural activity in detail in Chapter 3. Briefly, I postulated that the activation of a specific region yields from the computing and comparing of neural differences stemming from other regions. These regions, which serve to yield and amplify neural differences, may by themselves either be activated or non-activated. This means that even non-activated or non-recruited regions participate in generating neural differences.
Conceptually, one may therefore want to distinguish between “activated” and “active” regions. “Activated regions” are those regions that show neural activity changes in response to the task we apply. We as observers associate the recruitment of these regions with the function in question and are consequently inclined to localize the latter in the former.
This, however, neglects what I describe as “active regions” that do not show changes in their activity level in response to the task. These regions may nevertheless participate in generating the neural activity changes of the activated regions, more specifically in generating and amplifying neural differences (see what I describe as an “amplification hypothesis” in Chapter 3). They are thus “active” but not “activated.” This, however, makes localization of the function in the activated regions impossible, since that would neglect the role of the active regions in generating the neural activity changes in the activated region.
As detailed in Chapter 3, the generation and amplification of neural differences is coupled to the condensation of neural activity (see what I describe as a “condensation hypothesis” in Chapter 3). Rather than each of the original (p.312) lower order sensory regions’ activating a separate higher order cognitive region, the former’s neural activity converges in one common region, to which we then attribute localization. This, however, is a false-negative inference that focuses only on the outcome of localization in the higher-order cognitive region, while neglecting its underlying processes in which lower-order sensory regions participate.
More specifically, the outcome of sparse coding does indeed pertain to one particular region, the “activated” region or network as distinguished from all “non-activated” regions/networks. However, the underlying process involves “active” regions/networks (as distinguished from non-active regions/networks) that are essential in yielding and amplifying neural differences.
This means that the function in question cannot be localized exclusively and completely in the “activated” region/network itself. Instead, the function may be associated with both “activated” and “active” regions/networks as distinguished from “non-activated” and “non-active” ones. Accordingly, the regions/networks remaining silent in response to our task, that is, “non-activated, ” may nevertheless be “active” (rather than “non-active”) and may therefore have an important role in processing the function in question (see Fig. A2-1; see also Hayes et al. 2013 for an example of where the density of GABA-A receptors [PET] in ventromedial prefrontal cortex, a non-activated region during an aversive task, modulates the degree of signal changes [fMRI] in an activated region, the sensorimotor cortex; see also Gonzales-Castillo et al.
In sum, one may postulate localization of neural activity in specific “activated” regions/networks during particular function. This, as demonstrated, is the outcome of the processes guiding sparse coding on a regional level. However, to infer from such localization (or better, condensation) of neural activity to the localization of the function in question in that particular region/network is to confuse outcome and processes. The function in question must also be associated with regions (and networks) other than the “activated” ones like those that I here described as “active”; that is, actively involved in amplifying neural differences.
Neurotheoretical Remark IC: Complementarity between Holism and Localizationism
We are confronted with two apparently contradicting observations. On one hand, many regions, and ultimately, the whole brain, seem to be implicated in the neural processing of various psychological functions (see earlier). This suggests holism holds true on a psychological level. On the other hand, there is regional sparseness in that different psychological functions seem to recruit similar but at least strongly overlapping regions and networks. This observation, however, contradicts holism and would rather be compatible with localizationism holding true on a neuronal level.
How can we reconcile the contradictory assumptions of localizationism on the neuronal level and holism on the psychological level? The need to reconcile localizationism and holism was already recognized by K. Lashley, as is apparent in the following passage:
The chief advantage of the strict theories of localization has been their definiteness and comprehensibility. Those of us who have felt the inadequacy of such theories have had to fall back upon expressions like mass action, stress patterns, dynamic effects, melodies of movement, vigilance or nervous energy; all metaphorical and highly unproductive of experimental problems. Yet the facts demand something of this sort. The evidence seems conclusive that in various cortical functions there is every degree of specialization from a limited point-to-point correspondence of cells to a condition of absolute non-specificity. Not only is there diversity in the modes of action of different parts of the cortex but a single area, highly specialized and differentiated for one activity may be wholly undifferentiated for another in which it also participates. We have not a choice between a theory of localization and a theory of decentralization, but must develop a wider view which recognizes the importance and interdependence of both modes of integration. (Lashley 1931, 254)
I hypothesize that we need to set the alternative of localizationism versus holism into the context of sparse coding and difference-based coding in order to reconcile both. There is holism on the process level. As described earlier, even presumably silent, that is, “non-activated” but “active, ” regions/networks are nevertheless actively participating in generating and amplifying neural differences, thus allowing for difference-based coding. Such difference-based coding is in turn central in condensing and thus sparsening neural activity in a particular region, the “activated” or recruited region, as the manifestation of sparse coding on a regional level.
Many regions, if not (indirectly via the constitution of differences) the whole brain, actively participate in constituting neural differences. One may consequently suggest holism on the level of neuronal processes, whereas the very same neuronal processes, operating throughout different regions, allow and, even stronger, predispose the temporal and spatial sparsening of subsequent neural activity changes in a few highly localized regions as their outcome. The outcome, that is, the changes in the neural activity in a few localized regions, may thus be more localized when compared to the rather holistically operating processes.
What does this entail for the relationship between localizationism and holism? This means that the concepts of localizationism and holism are not opposite and contradictory to each other but rather mutually dependent on each other: As (p.314) there would be no outcome without a preceding process, localizationism would remain impossible without holism.
Even stronger, the more holistically processes that allow for the amplification of neural differences throughout the whole brain make a more localized outcome, that is, spatial and temporal sparsening of neural activity and the number of “activated” regions, almost necessary. Accordingly, localizationism and holism are bound together as tightly as process and outcome; they remain consequently as inseparable and complementary as yin and yang in the Chinese tradition.