How does Hemispheric Specialization contribute to Human-Defining Cognition?

Summary

Uniquely human cognitive faculties arise from flexible interplay between specific local neural modules, with hemispheric asymmetries in functional specialization. Here, we discuss how these computational design principles provide a scaffold that enables some of the most advanced cognitive operations, such as semantic understanding of world structure, logical reasoning, and communication via language. We draw parallels to dual processing theories of cognition by placing a focus on Kahneman’s System 1 and System 2. We propose integration of these ideas with the global workspace theory to explain dynamic relay of information products between both systems. Deepening the current understanding of how neurocognitive asymmetry makes humans special can ignite the next wave of neuroscience-inspired artificial intelligence.

Computational design principles behind the right and left hemisphere probably provide a scaffold that serves several human-defining cognitive feats. Hartwigsen and colleagues reason that better understanding this biological asymmetry for dual processing could ignite new forms of neuroscience-inspired artificial intelligence.

Introduction

The human body plan appears symmetrical from the outside only. We consistently lead with one foot when we climb stairs to enter a building. We fixate with our dominant eye when we span the bow and aim at the target in archery. We also naturally turn our right ear to a colleague who is whispering a secret message at a party. Understanding the meaning of the message works better through the right ear, given its privileged access to the language-dominant left hemisphere in the brain (the so-called right-ear advantage, Westerhausen, 2019).

Most biological systems show some degree of asymmetry (Geschwind and Galaburda, 1985). Functional asymmetries shape the human brain and behavioral functions in profound ways. Prototypical classes of human-defining cognition feature hemispheric differentiation in human brain organization. In particular, the left hemisphere plays a key role in communicative language functions and logical reasoning. Instead, the right hemisphere is relatively more specialized for spatial reasoning and emotional processing (Sun and Walsh, 2006; see below for details). Yet, neuroscientists are only beginning to understand hemispheric brain asymmetries. The precise adaptive advantages reaped from hemispheric asymmetries for cognition and behavior are still far from clear (Esteves et al., 2020).

Here we adopt a neurocognitive perspective on the computational design principles that may be a precondition for the emergence of human-defining cognition. The aim of our review is to point out salient parallels between asymmetries in i) biological neural networks, ii) special human cognitive functions, and iii) artificial neural networks. We recontextualize asymmetries in human neurocognitive functions and artificial neural network algorithms by integrating the System 1 – System 2 perspective. We argue that hemispheric asymmetries may provide a scaffold for information processing that evolved to enable an increasing degree of cognitive sophistication. Hemispheric specialization allows for parallel processing of several complex mental operations, like language and social cognition, that are uniquely powerful in the human species. We will mainly focus on language here because it is the key faculty for human communication. We will argue that insights into functional brain asymmetries could turn into a well of fresh ideas for designing the next breed of deep learning algorithms.

Hemispheric asymmetry as a core design principle of the brain

As a rule of thumb, we typically have one dominant hemisphere for certain types of cognitive operations. Such specialized responsibilities become especially apparent with regards to emotional responsiveness, visuo-spatial attention, as well as conscious problem solving and language capacities (cf. below). These functional asymmetries beg the question of the special purpose and neurocomputational infrastructure of how humans have evolved to perceive, understand, and navigate the world.

The emergence of the symmetrical body plan in animals likely co-occurred with improved mobility, more astute resource-seeking behavior and more effective predator defense (see Corballis, 2019). Why have human brains evolved to be organized into two specialized hemispheres? In the symmetrical brain, modular organization of specialized subsystems probably allows for a more efficient brain architecture, that is, denser packing of neurons in the same space (Kaas, 2000; Kaas and Herculano-Houzel, 2017; Van Essen, 1997). Brain organization into two hemispheres expands the amount of cortical surface area. This brain tissue growth was argued to afford advantages in terms of cooling and energy consumption (Kaas and Herculano-Houzel, 2017). Functional cerebral asymmetry expanded to override a purely symmetrical design of brain and body, and likely confers important evolutionary benefits.

The continuous refinement of specialized neural circuits may cohere with selection pressure for lateralization to enable more efficient functional organization and less redundancy between brain regions (Corballis, 2019; Vallortigara and Rogers, 2005). Moreover, hemispherically divergent processing is assumed to increase the ability to perform multiple tasks at the same time (Vallortigara et al., 2011), and prevent response competition between both hemispheres (Bisazza et al., 1998; Cantalupo et al., 1995). Having two exactly equivalent hemispheres would require numerous long-range connections in large-brained mammals such as human and non-human primates (Kaas, 2000). Long-range wiring comes at the cost of metabolic energy requirements for maintenance and conduction (Laughlin and Sejnowski, 2003; Lennie, 2003; Levy and Baxter, 1996; Ringo et al., 1994).

The emergence of unilateral functional networks might also have been favored by time constraints in information transfer via the corpus callosum between both hemispheres (Toga and Thompson, 2003). The necessary number of long-range axonal connections can be reduced by departing from the basic plan of symmetry in both hemispheres wherever possible, as well as grouping functions in the same hemisphere (Gazzaniga, 1995; Kaas, 2000). This need for local functional specialization (Sporns et al., 2004; Sporns et al., 2007) within each hemisphere may have compounded across primate evolution. This is because our monkey ancestors had to adapt to solve increasingly complex problems, such as coping with life in social groups (Kaas, 2000; see also Dunbar and Shultz, 2007; Tomasello et al., 2005). Of relevance for the following considerations, unilateral aggregation of function manifests in hemispheric asymmetry, thereby increasing the total neural capacity (Vallortigara and Rogers, 2005). Thus, asymmetrical features of functional brain organization enable efficient specialized and parallel processing of neurocognitive operations.

Functional specialization within each hemisphere is not meant to imply that each module works in complete isolation. Efficient processing in the asymmetrically specialized brain requires effective information transfer between both hemispheres (Davis and Cabeza, 2015; Genc et al., 2011; Westerhausen et al., 2011). Some authors argue that the need for flexible coordination of interhemispheric interaction is a direct consequence of brain lateralization (Hinkley et al., 2016; Karolis et al., 2019; Westerhausen et al., 2006). The development of the biggest white-matter pathway that channels information between both hemispheres, the corpus callosum (Schmahmann and Pandya, 2006), shortens the connection between both brain sides and the signal transmission times (Kaas and Herculano-Houzel, 2017; Paul et al., 2007).

As noted, long-range white-matter tracts come at the cost of increased metabolic energy demands for maintenance and conduction (Laughlin and Sejnowski, 2003; Lennie, 2003; Levy and Baxter, 1996; Ringo et al., 1994). The strength of connection through the corpus callosum is dependent on the extent of lateralization of functionally specialized regions. Homologous brain regions that are functionally distinct, that is, lateralized, tend to be less strongly linked to one another through the corpus callosum on a neural level compared to homologous brain regions that perform the same task (Karolis et al., 2019). This bottleneck constraint illustrates a potential evolutionary driver towards lateralization as a mechanism of increasing metabolic efficiency.

It is a repeatedly proposed idea that hemispheric asymmetries themselves may be a unique feature that characterizes the human brain. However, more recent accounts emphasize the presence of asymmetrical brain specialization in several distinct animal taxa (e.g., Corballis, 2009; Gunturkun and Ocklenburg, 2017; Vallortigara et al., 2011). Rodents, songbirds, and non-human primates show certain indicators of hemispheric asymmetry of brain function. For example, several avian species show prominent lateralization of visual processing (Gunturkun and Ocklenburg, 2017). The left hemisphere excels at categorization, discrimination, and memorization of visual patterns in different kinds of birds (Rogers, 2014; Valenti et al., 2003; Yamazaki et al., 2007). In contrast, the right hemisphere tends to assist more in visually guided interactions with emotional stimuli (Vallortigara et al., 2011; Ventolini et al., 2005), attentional shifts (Diekamp et al., 2005), social interactions (Vallortigara and Andrew, 1994), or relational analysis of pictorial and spatial information (Vallortigara et al., 2004; Yamazaki et al., 2007).

Some of these functional asymmetries appear to match those reported in human brains (see next section). Hemispheric asymmetries hence suggest themselves as a conserved design principle according to which localized functional specialization in the brain may bring a computational advantage. In fact, forms of hemispheric specialization have been identified in many species that evolved and live in different ecosystems. Hence, brain asymmetry probably emerged repeatedly and independently to boost cerebral processing capacity in a way that ultimately enhances survival (Gunturkun et al., 2020; Vallortigara and Rogers, 2005). Hemispheric specialization allows to improve execution of several differentiated processing operations in parallel. For example, when searching for food, it is crucial for survival to simultaneously screen the environment for predators and other threats (Gunturkun and Ocklenburg, 2017).

Given current knowledge, asymmetries in homologous brain regions in monkeys bear some degree of resemblance to left-hemispheric language-related brain regions in humans (e.g., Gannon et al., 1998; Hopkins et al., 1998; Sherwood et al., 2003). This observation suggests that brain regions linked to serving communicative abilities may take root before the inception of human language. Nevertheless, the functional relevance of these asymmetries in non-human primates remains unresolved (Vallortigara and Rogers, 2005). Based on the above-described parallels in hemispheric asymmetries between various animals, the more pertinent question is probably about the extent to which human brain lateralization is unique. That is, brain lateralization may be more appropriately viewed of degree, not of kind.

How does human brain organization then depart from that of non-human primates? A step-up in the cognitive sophistication and computational processing budget in the human species is closely tied to the considerable expansion of the higher association cortex in general as well as specific parts of the prefrontal, temporal and parietal cortex in particular (e.g., Glasser et al., 2014; Hill et al., 2010; Schleicher et al., 1999; Van Essen and Dierker, 2007). Asymmetries observed in comprehensive comparative assessments provide strong evidence that favor phylogenetic origins of brain lateralization (Toga and Thompson, 2003). Such clues show that differences between species are apparent especially in the inferior parietal lobule, aside from homologies between non-human primates and humans in the organization of certain brain areas (see Orban, 2016 for a review). Of relevance to our present considerations, there are probably 1.3 times more specialized cortical areas in humans than in monkeys.

Additionally, putatively unique human areas in the inferior parietal lobule are major contenders to make essential contributions to complex human-defining abilities, such as tool use or communicative language functions, problem solving, and future planning (Orban, 2016; Dohmatob et al., 2020; see also Seghier, 2013). There is also accumulating evidence for organizational elaboration of the prefrontal cortex between humans and other primates (Semendeferi et al., 2011; Teffer and Semendeferi, 2012). For example, prefrontal white-matter connections are consistently much denser in humans relative to other primates (Schoenemann et al., 2005). Grey- and white-matter increases of the prefrontal cortex may have played decisive roles in human cognitive evolution, which led up to the planning, language, attention and social information processing capacities that we have today (Hofman, 2014; Dohmatob et al., 2020).

In light of this evidence at the functional and structural level, the evolutionary trajectory of the human species has probably culminated in the most prominent hemispheric specialization in the entire animal kingdom (Gazzaniga, 2000; Vallortigara and Rogers, 2005). Similarly, the advancement of abilities for complex reasoning, such as required in recursive mental perspective-taking, are distinctly enabled by the way the human brain is organized for information processing. We speculate that a potential fitness gain obtained from stronger lateralization is asymmetrical specialization of neural circuits for enhanced dual processing. Extended functional specialization of one hemisphere to preferentially support a particular class of cognitive functions may maximize efficient, simultaneous information processing. Importantly, this computational design principle can also alleviate redundant computational efforts. More nuanced hemispheric specialization may also enhance advanced neurocognitive operations that depend on simultaneous processing of complementary information flows across both sides of the brain.

As an intermediate standpoint, we are the only species that has given rise to civilization and culture fueled by communication via language (Tomasello, 1999). This human-defining faculty shapes our complex everyday interactions with others and our society at large (Gelman and Roberts, 2017). For example, the ability to read and write books to pass on externalized knowledge to the next generation is a distinctively developed ability in humans, which sets us apart from the behavioral repertoire of other animals. As another example of this uniquely human faculty, the language-dominant left hemisphere has been attributed with integration of the distinct brain modules to form our unified ‘conscious awareness’, including the sense of being the responsible agent of our own actions (Gazzaniga, 2000). The way humans have evolved to solve problems about the world by deliberately hypothesizing, elaborating, and interpreting causal explanations of events may have unique roots in the left hemisphere (Gazzaniga, 2000; Dohmatob et al., 2020).

In the following sections, we discuss key computational strategies of the left and right hemisphere, with their relationships to domain-specific human-defining and domain-general processing in biological and artificial neural networks. We will outline that functionally specialized modules tend to be lateralized for increased overall processing efficiency, and human brains have a lot of such modules.

Asymmetries in biological neural networks

Some aspects of brain asymmetry, including anatomical features of language-related regions in the perisylvian cortex, become apparent already in the fetus in utero (Chi et al., 1977; Hering-Hanit et al., 2001). These observations speak in favor of genetic-developmental programs that are inherently lateralized, and again highlights asymmetry as a fundamental design principle of human brain organization (Kong et al., 2018). Aside from genetic influences, hemispheric asymmetries are generally thought to be shaped by evolutionary, developmental, environmental and in some instances also pathological factors (Toga and Thompson, 2003).

Structural and functional changes in brain organization during childhood provide examples of how developmental factors tie into the gradual maturation of hemispheric asymmetries. During cognitive development in childhood, structural changes of callosal connections, which reflect increased myelination and increasing overall size, allow for optimized neuronal communication between the left and right hemisphere (Danielsen et al., 2020). Such infrastructural refinements are likely related to development of lateralized language function. Structural brain changes may also support the development of related neural modules, which contribute to advanced cognitive capacities that draw on semantic world knowledge (Putnam et al., 2008; Thiel et al., 2006; Westerhausen et al., 2011). With respect to functional changes in hemispheric asymmetries, neuroimaging studies demonstrated that language is symmetrically organized in very young children (ages 4–6). However, the contribution of the right hemisphere was reported to decrease through childhood (Olulade et al., 2020; see also Basser, 1962; Lenneberg, 1967). Early hemispheric symmetry for realizing language-related capacities may reflect the plastic potential for reorganization in case of brain damage (Newport et al., 2017). Indeed, language may be equally affected by left or right perinatal stroke (Lansing et al., 2004). The same argument may explain gradual rather than absolute lateralization for most functions in the adult brain. This notion is also plausible as complete lateralization would increase the vulnerability to lesion-induced deficits and preclude reorganization by plastic recruitment of the non-dominant hemisphere (Thulborn et al., 1999; see Esteves et al., 2021 for a discussion). With respect to the role of environmental factors, active exposure to auditory stimuli facilitates the differentiation of the sounds in the environment during language learning (Benasich et al., 2014; Kuhl, 2004; Werker and Hensch, 2015).

Asymmetries in functional specialization

In general, prototypical kinds of cognitive operations feature hemispheric differentiation in brain organization, although this view is still controversial for some cognitive domains (Turner et al., 2015). The left hemisphere serves a crucial role in communicative speech and language capacities (Friederici, 2011, 2017; Hagoort and Indefrey, 2014), as well as mathematical and logical reasoning (Ashcraft et al., 1992; Goel, 2019). Instead, the right hemisphere is more functionally tuned to attentional reorienting operations (Corbetta and Shulman, 2002), such as in emotional processing (Demaree et al., 2005) as well as affective prosody (Belyk and Brown, 2014) and face processing (Prete et al., 2015; Vakil and Liberman, 2016). Episodic memory processing also features differences in hemispheric specialization, with the left hemisphere being dominant for encoding and the right hemisphere for retrieval (Habib et al., 2003; Johansson et al., 2020; Tulving et al., 1994).

Despite such functional asymmetries, both hemispheres closely interact during many cognitive operations. For example, sophisticated cognitive processes such as language comprehension and logic-based reflection that rely on the contribution of specialized circuits in the left hemisphere (Friederici, 2017; Goel, 2019) typically depend on simultaneous engagement of domain-general cognitive support functions (e.g., Blank and Fedorenko, 2017; Fedorenko et al., 2013). Domain-general support functions such as attention, cognitive control or working memory engage distributed neural networks in both hemispheres (see Asymmetries in human cognition for details).

Reading a book or newspaper requires left-dominant language competence, as well as simultaneous attentional routing that is preferentially supported by the right hemisphere. Different reading strategies require different interactions between these modularized neural systems: when we search for a specific keyword, we may need to ignore and suppress detailed sentence information, requiring inhibition of specific context. In contrast, when we are contemplating the logic of a particular paragraph, we may need to read every single sentence consecutively.

Notably, lateralization of a specific function does not imply that the respective other hemisphere is completely lacking specific processing abilities for that function. Indeed, asymmetries are normally a question of degree rather than strictly binary (Badzakova-Trajkov et al., 2016; Bartolomeo and Thiebaut de Schotten, 2016). As a prominent example, attention is commonly believed to be distributed across both hemifields in the intact human brain. However, the right inferior parietal lobule typically dominates over its left counterpart during bilateral attentional processes (Cicek et al., 2007). Accordingly, healthy participants usually show an attentional bias towards the left hemifield, a phenomenon called pseudoneglect (Bowers and Heilman, 1980; Zago et al., 2017). In neurological patients, this observation is consistent with a greater severity of attentional impairments after right relative to left parietal damage (Weintraub and Mesulam, 1987; see next section for details).

As such, not all cognitive functions are necessarily lateralized. For example, decision making tends to engage the prefrontal cortex in both hemispheres (Domenech and Koechlin, 2015). As discussed below, more domain-general functions appear to feature more bilateral recruitment. Indeed, neuroimaging studies provide evidence for distinct sub-specialization within the larger prefrontal cortex during decision making. The posterior medial part was associated with domain-general aspects of decision making and cognitive control, including conflict monitoring, error detection and confidence representation across a variety of tasks (Botvinick et al., 2004; Gehring et al., 1993; Ridderinkhof et al., 2004). In contrast, the anterior prefrontal cortex was related to tracking the reliability of specific alternative strategies during decision making (e.g., Donoso et al., 2014; Dohmatob et al., 2020). Recent work points to the coexistence of domain-agnostic and domain-specific representations of confidence in neighboring prefrontal subregions (Morales et al., 2018).

Asymmetries in structural organization

Primacy of functional dedication in both hemispheres in the human cerebral cortex is probably grounded in characteristic divergence of structural properties. Although both hemispheres are similar in total weight and volume, the spatial distribution of tissue differs markedly in the left and right hemisphere (Toga and Thompson, 2003). Considering structural differentiation in the entire brain, perisylvian regions, extending into the inferior parietal lobule, have shown the most prominent asymmetries, with larger volumes in the left hemisphere (Amunts et al., 1999; Falzi et al., 1982; Steinmetz, 1996). This feature of the human brain is likely related to the left-hemispheric dominance for language. Likewise, left-biased asymmetry in fiber density has been demonstrated for the arcuate fasciculus. This major white-matter pathway is known to connect language-related regions in the human brain (Buchel et al., 2004; Catani et al., 2007; Lebel and Beaulieu, 2009; Takao et al., 2011).

Functional asymmetries depend on specific contextual demands

It is one of the most established insights from neuroscience research that the left hemisphere harbors the superior language processor (Broca, 1865; Wernicke, 1874). Yet, this notion does not imply that the right hemisphere is invariably word-blind (Lindell, 2006). Indeed, meta-linguistic processing aspects such as metaphor processing or the processing of subordinate meanings of ambiguous words rely on a stronger contribution of right-hemispheric regions (see Hartwigsen and Siebner, 2012). In some cases, lateralization depends on the specific function of a given linguistic process.

This insight is epitomized by prosody, the change in rhythm and melody of speech (Cutler et al., 1997). Emotional prosody usually shows strong lateralization to the right hemisphere (e.g., Heilman et al., 1984; Hoekert et al., 2010; van Rijn et al., 2005). Likewise, the right hemisphere has been found to be involved in speech processing depending on the presence of pitch information (Meyer et al., 2002; Meyer et al., 2004). In contrast, linguistic prosody tends to be lateralized to the left hemisphere (Friederici and Alter, 2004; van der Burght et al., 2019; van Lancker, 1980; Belyk and Brown, 2014 for a meta-analysis). For example, functional neuroimaging work has demonstrated increased activity in the left prefrontal cortex when prosodic cues guide sentence comprehension (van der Burght et al., 2019). In contrast, when prosodic cues are superfluous for establishing the sentence structure, neural activity responses were lateralized to the right prefrontal cortex. Moreover, comprehension of meaningful words strongly engages the left hemisphere, while non-verbal semantic processing of meaningful environmental sounds depends especially on the right hemisphere (e.g., Butler et al., 2009; Thierry et al., 2003). Collectively, these results show that hemispheric asymmetries in the processing of meaningful stimuli strongly depend on the amount of linguistic information and its relevance for guiding language comprehension. This is an example of how flexible recruitment of specialized modules in our brain helps to master complex cognitive operations which depend on both, hemispheric specialization, and interhemispheric interactions. Indeed, some authors have argued that rapid excitatory and inhibitory interactions between neural circuits in both hemispheres are necessary for realizing particularly complex neural operations (Carson, 2020; Gunturkun and Ocklenburg, 2017; O’Donnell et al., 2017).

Probing the functional relevance of hemispheric specialization

How can we get a handle on the functional relevance of hemispheric specialization in humans? One opportunity is to study patients with circumscribed brain lesions in cognitive key regions. The anterior inferior parietal lobule is particularly pregnant to reveal new insights in brain lateralization. This is because this region is a hotspot of asymmetry in the human cortical mantle (Toga and Thompson, 2003).

Stroke-induced brain lesions in the left temporo-parietal cortex typically lead to impairments in language comprehension, semantic reflection and communicative abilities – known as Wernicke’s aphasia (Geschwind, 1970; Lichtheim, 1885; Wernicke, 1874; see Fridriksson et al., 2018). These neurological patients are usually able to speak fluently. However, circumscribed brain lesions around the temporo-parietal Sylvian fissure typically cause difficulties in the comprehension of spoken or written language. Patients may be unable to verbally describe pictures which require complex scene understanding. Depending on the severity of the neurological disorder, some patients may even be unable to articulate semantic concepts that occur in simple pictures. The semantic impairments substantially affect the abilities for interpersonal communication and social interactions in patients with Wernicke’s aphasia.

In contrast to impairments of human language and reasoning after brain lesions in the left hemisphere, selective tissue damage to the right anterior inferior parietal lobule is typically associated with specific impairments in visuospatial awareness and attentional routing processes – called hemispatial neglect (Corbetta and Shulman, 2011; Vallar and Calzolari, 2018). This neuronal damage causes the inability to process environmental information in the left visual field. Neurological patients with neglect typically behave as if the left side of their visual space did not exist. Hence, such individuals may ignore the left side of their body when taking a shower, or may shave only the right side of their face. While walking, neglect patients often collide with objects on their left side. Further, these patients may ignore and leave untouched food on the left side of the plate, as another example of the affected visuo-spatial capacities that are a consequence of damage to the right hemisphere. The proposed double dissociation in the functional relevance of the left hemisphere for language and the right hemisphere for visuospatial awareness and attentional reorientation receives support from a recent multi-site cohort study in >1,000 patients the early phase after stroke (Bonkoff et al., 2021; Figure 1).

Figure 1. Brain damage has hemisphere-specific consequences on specific cognitive capacities in 1080 stroke patients.

Causal evidence that circumscribed tissue damage leads to selective cognitive impairments in visuospatial function (RC, Rey Complex Figure Test), global cognition including space and time (MMSE, Korean Mini-Mental State Examination), language performance (BN, Boston Naming Test), and memory function (SVL, Seoul Verbal Learning Test). To achieve these single-patient predictions of clinical outcomes, the Bayesian hierarchical model relied on information of brain segmentation of lost grey-matter as input features. A) Heatmaps indicate the associations of each anatomical region (Harvard-Oxford atlas) with lost (blue color) or preserved (red color) cognitive performance in patients for different clinical assessments. B) Predictive accuracy (coefficient of determination R² estimated via posterior predictive checks) for the four cognitive scores along with scatterplots of actual (x axis) and predicted (y axis) cognitive performance. C) Brain renderings of the brain region-wise predictive relevance. Coordinates of axial brain slices in Montreal Neurological Institute (MNI) reference space. The population-level evidence shows that language-related and attentional-routing-related cognitive capacities are biased to the left versus right brain side. Reprinted with permission from Bonkhoff et al., Brain Comms, in press.

A central clinical question with respect to functional specialization is how the brain circuitry gradually adapts to recover from stroke and other tissue lesions in either hemisphere. We propose that functional specialization within each hemisphere and functional competition between both hemispheres shapes compensation when recovering from brain damage. Following tissue damage to the language network in the left hemisphere, recruitment of ipsilesional regions is usually associated with better recovery of function than recruitment of homologous regions in the right hemisphere (Winhuisen et al., 2005, see Hartwigsen and Saur, 2019; Saur and Hartwigsen, 2012 for review). The role of the non-dominant hemisphere in stroke recovery remains debated (e.g., Hartwigsen and Volz, 2021; Turkeltaub et al., 2011; Volz and Grefkes, 2016). Some studies in the language domain argue that right-hemispheric regions may be integrated, especially in case of large-scale damage to the language-dominant left hemisphere (Crosson et al., 2007; Heiss and Thiel, 2006). Indeed, longitudinal neuroimaging work shows that stronger recruitment of right prefrontal regions is associated with better language recovery (Saur et al., 2006; Stockert et al., 2020), at least in the so-called critical period (Krakauer, 2015) early after stroke.

How does the non-dominant hemisphere contribute to language recovery after large brain lesions? Recent work suggests that increased activity in right prefrontal regions during stroke recovery likely reflects domain-general cognitive control processes rather than language specific operations (Brownsett et al., 2014). There is increasing evidence that stronger recruitment of domain-general networks is associated with better language recovery after stroke (Geranmayeh et al., 2014; Geranmayeh et al., 2017). Many complex cognitive processes such as semantic reasoning, problem solving and planning require recruitment of diverse domain-general processes to realize their operation (Blank and Fedorenko, 2017).

At the neural level, this is reflected in the recruitment of multiple distinct, interacting brain networks (Petersen and Sporns, 2015). The more complex or demanding a cognitive process, the more disparate the set of cognitive processes that are typically co-recruited by default (see Diachek et al., 2020; Fedorenko and Blank, 2020; Fedorenko et al., 2013). Yet, engagement of such domain-general processes is unlikely to fully compensate for loss of elaborate mental performances that hinge on hemispherically specialized brain circuits (Hartwigsen, 2018; Heiss and Thiel, 2006). Compensatory mechanisms between left-hemispheric and right-hemispheric functional networks likely represent an overarching principle of neural organization that may be observed across diverse cognitive domains, such as language, abstract problem solving, imagination and planning (cf. Dohmatob et al., 2020).

Aside from the study of patients with brain lesions, functional neuroimaging in healthy volunteers is one of the key techniques to investigate hemispheric asymmetries in cognition (Ocklenburg et al., 2020). For example, a recent functional neuroimaging study demonstrated task-specific hemispheric specialization within the human inferior parietal lobule (IPL) (Numssen et al., 2021). The anterior part of the left IPL was strongest associated with a semantics/language task, while the anterior part of the right IPL showed the strongest predictive relevance for attentional reorienting. In contrast, both posterior IPL subregions showed the strongest association with social cognitive processing in the same set of healthy volunteers (Figure 2). Distinct domain-specific neural response patterns were reflected in distinct coupling patterns of the respective subregions. This observation pointed to stronger unilateral coupling for the anterior language and attention related regions and more complex bilateral coupling patterns for the posterior regions.

Figure 2. Hemispheric asymmetries in the inferior parietal lobule (IPL) characterize the neural substrates of human-defining cognitive domains.

A) IPL parcellation extracted from neural activity responses across attention, semantics and social cognition tasks in the same sample of healthy volunteers during functional brain scanning. Both IPL regions were divided into an anterior and a posterior subregion. B) In these task experiments, the IPL region showed a functional triple dissociation: the right anterior subregion had the strongest predictive relevance for attentional reorienting. The left anterior subregion had the strongest predictive relevance for lexical decisions. In contrast, both posterior subregions showed the strongest relevance with a social cognitive task that requires contemplating other individuals’ mental states. C) Task-evoked functional connectivity profile for the social cognition task. Both posterior subregions showed spatially distributed connectivity patterns with brain-wide cortical networks. Left anterior IPL: L-ant, right anterior IPL: R-ant, left posterior IPL: L-post, right posterior IPL: R-post. Networks: DAN: dorsal attention network, DMN: default mode network, FPN: fronto-parietal network, LIM: limbic network, VAN: ventral attention network, VIS: visual network, SMN: somatomotor network. Adapted and reprinted with permission from Numssen et al., eLife 2021.

Collectively, these results support the idea of hemispheric asymmetries for language and attentional allocation. The bilateral recruitment of the IPL for social cognition converges with previous meta-analyses demonstrating that advanced social cognitive functions, such as the capacity to infer others’ thoughts, beliefs, and behavioral dispositions, engage the IPL in both hemispheres (Bzdok et al., 2016; Bzdok et al., 2013; Schurz et al., 2020). The contribution of both hemispheres to social cognition may reflect the increased complexity of social cognitive operations, engaging both semantically tuned processes in the left hemisphere (Bzdok et al., 2016) as well as affectively tuned processes in the right hemisphere (Bzdok et al., 2013).

Asymmetries in human cognition

We speculate that there may be a connection between aspects of hemispheric specialization for different cognitive functions (cf. above) and Kahneman’s System 1 and System 2 (Kahneman, 2011). System 1 describes fast modes of cognition such as automatic gut responses to seeing another person’s angry facial expression or being surprised by a sudden harsh noise, without much conscious access or the results of deeper reflection. According to Kahneman, such impressions and feelings drive much routinized behavior in everyday life. For example, we often trust our gut feeling when judging other people or deciding if we want to either further engage in a cooperation or distance ourselves. Kahneman argues that the majority of action-perception cycles throughout our days are mainly paced by System 1 cognition, maybe in a kind of autopilot. Important to our present considerations, sensory perception, reactive processing and mounting stereotypical motor responses are behaviors that humans share with many other animals (see Gunturkun and Ocklenburg, 2017).

It is the System 2 mode of thinking that drives the types of cognition that are unique to the human species: deliberate reasoning and detailed reconsideration to override prepotent emotional responses and actions, effortful rationalization, and manipulating representations of people’s world view (Hofman, 2015). Such cognitive operations may prevail when solving a creative planning problem such as “Which alternative street paths could I drive back home, given unexpected construction on the road?”. Contemplating candidate solutions to such problems usually takes more time than problems that can be solved exclusively with cognitive processes that belong to System 1. When concentrating on figuring out a challenging planning and navigation task, humans are less likely to be triggered or distracted by visual and auditory cues or sources of affective arousal from the environment (e.g., ‘Invisible Gorilla’ Test; Chabris and Simons, 2010).

System 2 cognition is thought to reject, alter, or overcome impressions, intuitions, feelings, and reactive tendencies that are issued by System 1 (Kahneman, 2011). System 2 can also fully endorse and select from the bottom-up inputs provided by System 1. The division of labor between the two cognitive modes likely optimizes everyday life in terms of allocated resources and spent cognitive effort. Both System 1 and 2 probably work hand-in-hand and are permanently active in degrees. System 1 may be sufficient for handling most routine decisions. System 2 may come into play as we encounter unusual objects, new situations and other cognitively challenging tasks.

To efficiently execute advanced cognitive operations, System 2 routinely exerts tight governance over various System 1 processes. Instead, System 1 processes do not typically depend on System 2 operations. One example for the strong interaction between both systems would be mentally browsing candidate driving routes (outside of one’s visual environment). This operation requires recruitment of System 2 for coordination between System 1 modules, involving retrieval of structured world knowledge, visual imagination, planning, as well as attentional and working-memory processes. In contrast, executing familiar everyday actions, such as driving straight on an already chosen and known driving path do not require effortful reasoning or sophisticated semantic representations.

Notably, to the best of our knowledge, System 1 and System 2 have not been explicitly mapped onto specific brain regions (Kahneman, 2011; Evans and Stanovich, 2013; Kannengiesser and Gero, 2019). The exact mapping of particular cognitive processes to System 1 and System 2 cognition can also be challenging. For example, there are different facets of attentional routing mechanisms, including conscious and automatic forms of attention (Turatto et al., 2000). According to Kahneman, System 2 typically depends on conscious attention. Instead, System 1 (e.g., in visual object detection) is believed to operate with automatic or subliminal attention, without the need to deliberately focus on a specific aspect of one’s internal or external landscape (e.g., processing objects that occur outside of the currently focused field of view).

Can we link System 1 and System 2 to different specialized circuits in the human brain? We are tempted to draw some tentative parallels to hemispheric specialization. We speculate that System 1 may be reflected in localized processing of distinct regions or networks. Depending on the specific subserved process, these local neural systems may engage circuits in the left hemisphere, right hemisphere, or both conjointly. As cognitive operations grow more complex, they probably prompt recruitment of more effortful System 2 cognition. System 2 engagement, in turn, may involve recruitment of both specialized networks and domain-general support functions.

For example, language comprehension is known to invoke a collection of functionally specialized circuits that are biased to the left hemisphere (e.g., Binder et al., 2009; Binder et al., 1997; Fedorenko et al., 2012; see also discussion above). With rising task demands, language comprehension is increasingly supported by domain-general cognitive and executive control functions to enable efficient processing (Fedorenko et al., 2013; Duncan and Owen 2000). At the neural level, domain-general process engagement is reflected by increased recruitment of bilateral domain-general networks, especially the ‘multiple-demand network’ for cognitive control (Duncan, 2010). The multiple-demand network includes brain regions that respond to increasing task demands across many cognitive domains and is involved in the planning, top-down control, and regulation of cognition and goal-directed behavior (Dosenbach et al., 2006; Dosenbach et al., 2007). This network engagement was reported to scale as a function of task difficulty across a variety of cognitive operations (Duncan and Owen, 2000; Fedorenko et al., 2013). Previous work demonstrated that simple story listening, which closely matches natural language processing, mainly recruits the specialized left-lateralized language network (Blank and Fedorenko, 2017; Diachek et al., 2020). In contrast, more complex listening scenarios, such as listening to speech in noisy environments, strongly engage additional multiple-demand systems (e.g., Peelle, 2018; Vaden et al., 2013; Wild et al., 2012).

By analogy to Kahneman’s cognitive systems, listening to a story is a relatively automatic and routinized process which is probably mainly supported by System 1. In contrast, increased task demands under challenging listening conditions lead to higher cognitive load. We reason that such effortful, System 2-like task processing requires interaction of the specialized language network with the bilateral multiple-demand network in the example above. To be clear, we do not mean to argue that language equals System 1 processing, nor that increased cognitive control equals System 2 cognition. We rather suggest that System 1 cognition translates to all automatic, overlearned processes which may engage relatively local specialized networks. In contrast, System 2 cognition would engage effortful cognitive processes that rely on more distributed processing of collective networks and flexible interactions between these specialized neural systems (Figure 3).

Figure 3. Two examples of cognitive problem solving that engage preferentially System 1 (left) or also System 2 (right).

A) Finding the solution to this problem is easy and quick. System 1 answers this question without prompting in one second at most. Human observers do not necessarily know how to describe the average length in centimeters. However, we are accurate in adjusting the length of another line to match the average. Involvement of System 2 is not needed to form an impression of the norm of length for an array. System 1 achieves this task automatically and effortlessly, just as it recognizes the color of the lines and the fact that the lines are not parallel to each other. B) A glance provides an immediate impression of many features of the figure. We realize that the two towers are equally tall and that they are more similar to each other than the tower on the left is to the array of blocks in the middle. However, the observer does not immediately know that the number of blocks in the left-hand tower is the same as the number of blocks arrayed on the floor. To confirm that the numbers are the same, we would need to deliberately count the two sets of blocks and compare the results, an activity that requires recruitment of System 2. Adapted and reprinted from Kahneman, 2011.

However, it remains incompletely understood how information is efficiently relayed between System 1 and System 2 (Shleifer, 2012). Some authors proposed such information transfer between neural modules to be enabled by the global workspace (Dehaene et al., 1998; Mashour et al., 2020). In short, the global workspace theory was introduced to make steps towards explaining conscious processing (Baars, 1988). This conceptual framework was later extended to incorporate neural processes underlying effortful cognitive tasks (Dehaene et al., 1998). These authors distinguish two computational spaces: one global workspace composed of distributed and heavily interconnected neural circuits with long-range connections, and a set of specialized modules committed to perceptual, motor and cognitive processing (Dehaene et al., 1998; Mashour et al., 2020). Workspace neurons are mobilized in effortful tasks for which the specialized processors dissatisfy the needs to reach the processing goal. Thereby, they mobilize or suppress the contribution of specific processor neurons and enable conscious processing (see also Baars, 1988).

Supported by evidence from computational modeling, the global workspace appears to be central during the acquisition of novel tasks and effortful execution. The governance of the global workspace fades during highly overlearned tasks, but resumes sharply after erroneous task responses (Dehaene et al., 1998). As such, the global workspace satisfies key requirements for coordinating between System 1 and System 2: It is absent during routine tasks (requiring System 1 only) and takes importance when a novel non-routine task is introduced (requiring System 2). As an important prediction from this framework, the global workspace is recruiting distributed brain regions that have ample long-distance connections (Dehaene et al., 1998). The global workspace is thought to integrate multiple specialized brain regions in a coordinated and flexible manner during conscious processing. Thereby, the global workspace enables information transfer between specialized modules in both hemispheres, by integrating information across distributed neural processors (Mashour et al., 2020). Since distributed workspace neurons are mobilized in effortful tasks for which specialized processors do not suffice, we reason that workspace regions may overlap to some degree with domain-general regions of the multiple-demand network. Indeed, it was suggested that, among other regions, the dorsolateral prefrontal and anterior cingulate cortex contribute to the workspace (Dehaene et al., 1998). These regions have been associated with effortful task processing (Cohen et al., 1997; Pardo et al., 1990).

In short, the global workspace is thought to act as a central router that mediates intermediate processing outcomes to be amplified, sustained, and broadcasted to specialized neural modules that act as processing partners (Mashou et al., 2020).

Asymmetries in artificial neural networks

Various types of pattern-recognition tasks are solved by the human brain in less than one second – without effort, deliberate concentration or conceptual elaboration. We quickly detect other people’s emotional experiences and identity from parsing cues in their faces, mimics and body gestures (cf. Bzdok et al., 2012). We easily recognize a friend and her current affective state from the sound of her voice on the phone; without visual input from her facial or body expressions (cf. Hensel et al., 2015). We rely on automatic and fast memory retrieval to know whether we have heard a particular music melody before. This comes easy although we have been exposed to many thousands of different music pieces over our lifetime.

All these pattern-recognition computations can be carried out by the human brain in dozens or few hundreds of milliseconds (Dehaene et al., 2017). Notably, it is also these kinds of visual and auditory signal detection problems that have been revolutionized by deep learning algorithms over the last 10–15 years (Goodfellow et al., 2016). In some tasks, new deep learning architectures have even exceeded human-level performance (Ghahramani, 2015; Jordan and Mitchell, 2015; LeCun et al., 2015). In the terminology developed above, the artificial neural networks invented so far have proven effective in solving problems that are largely reminiscent of Kahneman’s System 1 cognition. Designing artificial neural network algorithms to take on tasks that involve Kahneman’s System 2 cognition remains more daunting. To tackle high-level problems, humans often require more than 1 second (Lake et al., 2017). For example, today’s deep learning architectures still struggle to understand the intentions of agents or causal implications in complex visual scenes based on images or photos, partly because of lacking modules that are specialized in representing explicit world knowledge (cf. Bengio, 2017; Marcus and Davis, 2019).

Existing deep learning systems excel at narrow tasks for which we have a sufficiently massive dataset of representative and correctly labeled instances or data points. However, if these artificial neural networks are trained to solve a task at hand using for example medical images from one hospital, their performance on the same kind of medical images from another hospital may be much worse (sometimes called ‘distribution drift’) (Tsymbal, 2004). This is due to changes in why, how, and when these medical images of human individuals are acquired. Such changing data distributions may be due to the practices or equipment that diverge between the two hospitals. The ability to generalize well or quickly out-of-distribution, from example observations freshly acquired from alternative sources, is much stronger in humans than in current artificial intelligence systems (Goyal and Bengio, 2020). This realization begs the question of how humans deal with newly encountered settings. Humans appear to resort to System 2 abilities (such as required for deliberate planning) to reason about such changes in the distribution in their environment and adjust their predictions accordingly. This partial robustness of human judgment to volatility in the world is stimulating deep learning research by inspiration from the cognitive neuroscience of System 2 abilities (Bengio, 2017; Goyal and Bengio, 2020).

Importantly, we do not argue that current artificial intelligence systems lack all aspects of what constitutes System 2 abilities. For example, the similarity question in Figure 3B represents a counting example in which today’s artificial intelligence systems have been reported to outperform humans. Rather, today’s artificial neural networks struggle in many more elaborate tasks, which require common-sense abstraction or generalization to new categories of encountered problems. For example, the task “play a song in C minor” necessitates System 2 thinking but would not be unfeasible for humans with musical training. In contrast, current artificial intelligence systems struggle with many such creative tasks.

For several decades, there has been vibrant crosstalk between the design of artificial intelligence systems and neuroscientific ideas about how the brain and its functionally dedicated modules work. Many such fruitful analogies were largely inspired by hierarchical processing cascades in the early sensory cortices of the primate brain. For example, principal component analysis and convolutional neural network algorithms applied to images of natural scenes do recover atomic features, such as edges and shapes. These image-derived features are indeed similar to current knowledge of the consecutive processing stages of environmental stimuli in the visual cortex (Hubel and Wiesel, 1962; Olshausen and Field, 1996; Yamins et al., 2014).

We emphasize that, up to now, the design of artificial neural network architectures has successfully mimicked several computational design principles that are not specific to human cognition, such as visuo-spatial object recognition, attentional selection, and raw memory retrieval. These influences from neuroscience on developments in the deep learning community are more naturally placed in Kahneman’s System 1 mode of cognition. As a logical next step, deep learning will probably thrive on new algorithms that capitalize on incorporating more advanced forms of human neurocognition and their functional interaction with brain-wide neural systems. New impulses may allow for a more complete understanding of how human-defining cognition is instantiated by brain circuits. Critical insights may be derived from the distributed representation across and functional asymmetry between two architecturally almost identical subsystems – the modular circuits from biological neural networks that together form the left and right brain hemisphere.

Stronger functional compartmentalization of artificial neural networks can yield many benefits, which are exemplified in the earlier idea of ‘capsules’ (Sabour et al., 2017). More generally, modular neural networks (Andreas et al., 2016; Goyal et al., 2019; Jacobs et al., 1991) are increasingly proposed that have modules that communicate not scalars but bundles of network activations. Building more functionally modularized infrastructure into the modeling architecture allowed learning more robust internal representations or concepts, with the ability to tackle novel settings by dynamically combining the pieces of knowledge contributed by different modules.

Decomposition of knowledge into independent mechanisms has been argued to be a key checkpoint on our path towards capturing causal structure with learning algorithms (Goyal and Bengio, 2020; Goyal et al., 2019; Peters et al., 2017). Distinct modules could implement such a specialization. The global workspace theory (Baars, 1988; Dehaene et al., 2017; see above) proposes that conscious processing involves the selective participation of only a few of the available expert modules at a time, then broadcasting information products to the rest of the brain through a communication bottleneck. Conscious attention selects which module and which conscious content is gated through this bottleneck and remains available shortly in working memory. In this way, with certain parallels to attention processes in brain and cognition (cf. above), computational attention mechanisms offer an algorithmic opportunity to negotiate between incoming bottom-up information and governing top-down modulation from deeper processing layers (Mittal et al., 2020). As argued recently (Bengio, 2017; Goyal and Bengio, 2020; Goyal et al., 2019), this dynamic selection of computational modules may bolster the ability to generalize in systematic and robust ways which the homogeneous and static architecture of more traditional deep learning systems does not enjoy (Mittal et al., 2020). In addition, this localization and factorization of knowledge into the appropriate modules can confer an advantage in terms of effective adaptation to causal understanding of and contingent action on the environment (Bengio et al., 2020).

The exchange of intermediate computing products between dedicated information processing modules can be controlled through an attention mechanism that “spotlights” pieces of information for active read-out. Coherent interpretation of input information can be achieved by algorithmic focus on a single or a small number of elements at once – a key ingredient that can be instrumental for realizing human-level artificial intelligence. Global-workspace components integrated into emerging deep neural network architectures get to decide where a given module reads information outputs from (Goyal et al., 2019). At a given time, only a subset of the functionally specialized modules, which compose the artificial neural network, are in sync with other parts of the modeling architecture (Figure 4). This mimics the realization from neuroscience research that performance gains in neural processing are obtained from imposing a bottleneck: any module cannot communicate with any other module all the time. Instead, purposeful scarcity of information flows is imposed by competition between expert modules to solve the down-stream learning goal. Indeed, human minds typically can hold only 7 +/−2 elements in working memory (Baars, 1988; Miller, 1956).

Figure 4. Flexible information processing in artificial neural networks through a global workspace.

Unconstrained pairwise communication between modules can be replaced by deliberately imposing a bottleneck on information communication via a shared workspace. Only a constrained subset of expert modules can “talk to” other expert modules at any given time (Mittal et al., 2020). Thus, only a few pieces of information are considered by the learning architecture in any given moment – a rough analogue to what is “consciously perceived”. Actively communicating expert modules are shown in blue. In a two-step process, expert modules first compete against each other for write permission to the shared workspace. Then, the ensuing processing outputs from a (purposefully restricted) module collaboration is made available to all modules of the artificial neural network architecture. The Figure shows example extensions of backbone architectures (top row) by equipping them with a global processing workspace (bottom row). RIM: Recurrent independent mechanisms, TIM: transformers with independent mechanisms, SW: shared workspace. Reproduced with permission from Goyal et al., 2021.

The collaboration of engaged functionally specialized modules is permitted to write into the shared workspace. The information, issued by the channeled module interaction, is then broadcasted for reuse by the entire modeling architecture. Every other module can collect any piece of information from that common buffer of intermediate processing outputs. The more parsimonious the set of dynamically interacting recurrent mechanisms, the bigger the advantage from compartmentalized information processing (Bengio, 2017). In this way, attention mechanisms can also assist in consulting and integrating declarative world knowledge from semantic expert modules. Their implication in processing cascades in artificial neural networks in turn may be a necessary step for realizing ‘elements of a conscious thought’ or for solving problems that ‘require a sequence of conscious steps’ (Bengio, 2017). Indeed, the election of pieces of information to be picked for global dissemination has been proposed to be a crucial aspect of how consciousness is instantiated in the human brain, and may be in machines (Dehaene et al., 2017). These advances in architecture design have led to more robust hidden representations and promising performance gains in a variety of challenging prediction tasks, especially those involving out-of-distribution generalization (Lake et al., 2017). Such brain-inspired enhancements of deep neural network computation can make strides towards human-defining reasoning, planning, and decision-making by algorithmic inference.

The negotiation between bottom-up processing of sensory inputs and top-down modulation calibrated by semantic world knowledge could make progress towards learning agents with improved performance at a variety of advanced tasks. Decomposable pieces of high-level semantic world knowledge can be flexibly induced, adjusted and recombined for binding with information products from other neural network modules – including not only visual object recognition but also attention mechanisms. Importantly, attending and composing the appropriate modules, and abstract entities they represent, has the potential for systematic generalization in non-stationarity environmental dynamics, that is, solidified out-of-distribution extrapolation (Goyal & Bengio, 2020). Such intersections between designated processing modules may make steps towards realizing aspects of causal reasoning, and more System 2-like deep learning systems that robustify generalization from one context to future contexts.

It has been argued that even the most recent deep learning systems remain limited to what corresponds to nonconscious processing in the biological neural networks of the human brain (Dehaene et al., 2017). Progress has recently been made in extremely large artificial neural networks for processing natural human language. These architectures were composed of ~175 billion model parameters, making them several orders of magnitude bigger than its predecessors (Mindt and Montemayor, 2020; Radford et al., 2018). Generative Pre-trained Transformer 3 (GPT-3) was released in mid-2020, after ingesting a diverse series of massive text corpora (as much as >40 TB). The trained learning system can produce human-like text in a variety of contexts, including ad-hoc conversation and completing the next sentence. In many cases, text created by GPT-3 is hard to tell apart from text that was written by humans. However, critics have flagged inaccurate world representation as a critical flaw in the GPT-3 architecture. The generated text passages were typically perfect in appearance. Yet, discrepancies repeatedly loomed in world understanding and common-sense logic, such as reflected in poor social, physical, biological, or causal reasoning (Marcus and Davis, 2020). Moreover, such state-of-the-art learning systems tend to have difficulty in connecting semantic information with long-term dependencies, such as following the narrative in longer stories (cf. Pereira et al., 2018). Hence, today’s language models mostly adhere to the brute-force agenda of interpolating between swaths of encountered language production patterns, rather than extrapolating principled rules from them (cf. Dreyfus, 1972).

Many authors have advocated that we need to go beyond learning superficial internal representations of happenstance word correlations. To this end, next-generation artificial intelligence systems may have to incorporate a special semantic or symbolic module that can dynamically liaise with other expert modules (cf. Chomsky, 1965; Fodor and Pylyshyn, 1988; Marcus and Davis, 2019; Pinker, 1994). A form of a-priori knowledge or set of priors can prove pivotal to go beyond narrow-sense artificial intelligence by discovering and manipulating strong candidate representations as a cornerstone to understand meaning in the external world (Chollet, 2019; Marcus, 2020; Marcus and Davis, 2019). In short, even today’s best-performing artificial intelligence systems have little grasp of and are untethered from reality; at least they are not cognizant of world structure in the way humans understand it. Hence, it may be no accident that it is precisely language understanding which is the human-defining capacity that displays such a strong hemispheric asymmetry between the left and right brain hemisphere.

Conclusions and Outlook

The human central nervous system may be the best existence proof for the feasibility of artificial general intelligence. This begs the question of the computational design principles of how processing modules of the human mind and brain are anchored in hemispheric asymmetry. After all, in nature and technology, intelligent systems often arise from flexible interaction between a small set of simple system components (Braitenberg, 1986; Fodor, 1983; Minsky, 1988). Such information processing systems may better capture pertinent structure in the world and provide a blueprint for more robust generalization to new environmental changes and challenges.

In particular, evolutionarily younger parts in the human inferior parietal lobule display some of the widest structural and functional divergences between the left and right hemisphere in the human brain. The functional role of the right IPL is tuned to serve attentional routing for information relay between specialized modules in our biological neural networks: this finds an algorithmic counterpart in attention-directing mechanisms that are increasingly boosting the capabilities of modern artificial neural network systems. In contrast, the functional role of the left IPL in language understanding, semantics-related reasoning and binding knowledge of world structure, which reflects System 2 thinking, is still awaiting an algorithmic counterpart in future developments that can propel deep learning.

Hence, thrusts forward in understanding the division of labor between the left and right brain hemisphere bear the potential to stimulate new innovations in deep learning. Emerging artificial intelligence tools may in turn unlock new neuroscience insights from big data, percolating into novel treatment interventions in various brain disorders with a disbalance in hemispheric asymmetry.