*4.2.2 Putting it all together: neuronal substrate of understanding and brain functional architecture*

Assumptions advanced in the preceding section entail the following suggestions.

1.Formation of neuronal packets transforms associative network into a packet network embedded into an energy landscape, with the packets residing in local minima. The height of packet energy barrier *Em* (free energy) is a function of temperature T and parameter σ(*T*) reflecting cumulative strength of

associative links incident to the packet's Markov Blanket (MB) from inside the packet vs. the cumulative strength of those incident from the outside.

$$E\_m = \sigma(T) - T \, d\sigma(T) / \, dT$$

#### **Figure 11.**

*Imagine raising your arm and touching the tip of your nose in three consecutive positions: looking to the left, looking straight ahead, and looking to the right. Population vector in the packet determining arm movement rotates accordingly. Coordinates of the nose tip in mental space control tuning of numerous neurons in the arm packet. Seminal studies in [51–53] demonstrated that movement organization involves rotation of packet vectors in the direction of the target.*

**(**σ (*T*) is analogous to membrane potential determined by the difference in ions concentration on both sides of the membrane, σ (*T*) declines as temperature grows, *Em* is an inverse of MB's permeability (resistance)). Packets connected by associative links might not be mutually accessible if separated by high energy barriers, as illustrated in **Figure 12**.

The height of energy barrier *Em* determines relative stability of packet *Xm* that corresponds, roughly, to a level of subjective confidence in *Xm* , which can vary depending on the local temperature (the lower the temperature, the higher the barrier. Consistent with [54], temperature variations shape the landscape and facilitate jumps of free energy barriers. Under the notion that deployment of neuronal resources serves to extract free energy from the environment [11, 12] temperate can be viewed as a control parameter regulating access to intra-packet resources, which equates temperature inverse to a cost, in entropy, of the free energy reward from the outside [54] received by the system as a result of the packet's deployment). The subjective experience of local temperature corresponds, roughly, to a level of arousal associated with object α *<sup>m</sup>* . As a result, circumstances are possible when packets having low evidential support (low cumulative strength of internal associations) remain stable, separate from other packets and inaccessible to coordination with them.

2.Variations in the mode of energy delivery (level of arousal, sustained and focused attention vs. wandering and diffuse attention) cause deformations in the landscape and enable overcoming energy barriers. **Figure 13** illustrates these notions.

Maintaining focused attention underlies the experience of cognitive effort that accompanies recall or attempts to ascertain connections between some entities (e.g., objects represented by packets *Xi* and ) *Xk* . The experience was best described in [56], as shown in **Figure 14**.

3.Mental models are synergistic neuronal complexes that comprise packets, regulatory neuronal structures that coordinate rotation of packet vectors, and

#### **Figure 12.**

*Here, q denotes a coordinate in the packet network space packets Xm and Xi are adjacent in the network but are not mutually accessible due to a high energy barrier that separates them. By contrast, packets Xi and Xk are mutually accessible (think of a terrain where Xi and Xk settlements are located in the same valley and are separated by a steep hill from Xm ).*

**91**

**Figure 14.**

**Figure 13.**

*Brain Functional Architecture and Human Understanding*

*cortico-thalamo–cortical connections, vs. cortico-cortical connections).*

excitatory-inhibitory connections between the packets serving to constrain

processes became decoupled from the motor-sensory feedback. The hypothesis is that neuronal machinery of sensory-motor coordination richly developed in the protohuman was adopted for the task of mental coordination not accompanied by any overt activities [28]. As a result, neuronal mechanisms could retain a rich repertoire of coordination capabilities but became unencumbered by the spatio-temporal constraints facing sensory-motor acts (e.g., when raising a hand, one cannot skip over intermediate positions or exceed the range and speed limits afforded by the muscular-skeletal system. By contrast, envision-

4.Mental modeling entered the stage (i.e., Sapience emerged) when mental

*The experience of mental effort. "Call the forgotten thing Z, the first facts with which we felt it was related to a, b, and c, and the details finally operative in calling it up 1, m, and n. The activity in Z will at first be a mere tension; but as the activities in a, b, and c little by little irradiate into l, m, and n … their combined irradiations upon Z succeed in helping the tension there to overcome the resistance, and in rousing Z to full activity. Through hovering of the attention in the neighborhood of the desired object, the accumulation of associates becomes so great that the combined tensions of their neural processes break through the bar, and the* 

*1) elevated arousal combined with diffuse attention equate to increasing temperature across patches in the packet network, causing temporary lowering of energy barriers and enabling inter-packet coordination (term 'cognition' derives from the Latin* cogitare*: Shaking together [55]. 2) sustained, focused attention equate to targeted energy delivery sufficient for local lowering and overcoming of the energy barriers, enabling coordination (term explanation derived from the Latin* explanare*: Flatten, make level or plane (Harper-Collins Dictionary of Philosophy, 1992). 3) inter-packet coordination can involve structures residing outside packet network (i.e.,* 

Decoupling from motor-sensory feedback created a gateway into mental universe populated by products of composition (imagination). To yield adaptive benefits, regulatory mechanisms were needed that would curtail superfluous compositions and facilitate those that could be mapped back onto and benefit overt behavior (i.e., allow predictions). Understanding is such a mechanism: although being rooted in sensory-motor coordination, understanding allows predictions unrestricted by spatio-temporal limitations of sensory-motor processes

vector rotation. **Figure 15** illustrates these notions.

*nervous wave pours into the tract, which has so long been awaiting its advent" ([56] p. 586).*

ing the same act does not face such restrictions).

*DOI: http://dx.doi.org/10.5772/intechopen.95594*

*Brain Functional Architecture and Human Understanding DOI: http://dx.doi.org/10.5772/intechopen.95594*

#### **Figure 13.**

*Connectivity and Functional Specialization in the Brain*

ions concentration on both sides of the membrane,

high energy barriers, as illustrated in **Figure 12**.

sponds, roughly, to a level of arousal associated with object

packets and inaccessible to coordination with them.

described in [56], as shown in **Figure 14**.

(*T*) is analogous to membrane potential determined by the difference in

The height of energy barrier *Em* determines relative stability of packet *Xm* that corresponds, roughly, to a level of subjective confidence in *Xm* , which can vary depending on the local temperature (the lower the temperature, the higher the barrier. Consistent with [54], temperature variations shape the landscape and facilitate jumps of free energy barriers. Under the notion that deployment of neuronal resources serves to extract free energy from the environment [11, 12] temperate can be viewed as a control parameter regulating access to intra-packet resources, which equates temperature inverse to a cost, in entropy, of the free energy reward from the outside [54] received by the system as a result of the packet's deployment). The subjective experience of local temperature corre-

circumstances are possible when packets having low evidential support (low cumulative strength of internal associations) remain stable, separate from other

2.Variations in the mode of energy delivery (level of arousal, sustained and focused attention vs. wandering and diffuse attention) cause deformations in the landscape and enable overcoming energy barriers. **Figure 13** illustrates these

3.Mental models are synergistic neuronal complexes that comprise packets, regulatory neuronal structures that coordinate rotation of packet vectors, and

*Here, q denotes a coordinate in the packet network space packets Xm and Xi are adjacent in the network but are not mutually accessible due to a high energy barrier that separates them. By contrast, packets Xi and Xk are mutually accessible (think of a terrain where Xi and Xk settlements are located in the same valley and are* 

Maintaining focused attention underlies the experience of cognitive effort that accompanies recall or attempts to ascertain connections between some entities (e.g., objects represented by packets *Xi* and ) *Xk* . The experience was best

ture grows, *Em* is an inverse of MB's permeability (resistance)). Packets connected by associative links might not be mutually accessible if separated by

σ

α

*<sup>m</sup>* . As a result,

(*T*) declines as tempera-

**(**σ

notions.

**90**

**Figure 12.**

*separated by a steep hill from Xm ).*

*1) elevated arousal combined with diffuse attention equate to increasing temperature across patches in the packet network, causing temporary lowering of energy barriers and enabling inter-packet coordination (term 'cognition' derives from the Latin* cogitare*: Shaking together [55]. 2) sustained, focused attention equate to targeted energy delivery sufficient for local lowering and overcoming of the energy barriers, enabling coordination (term explanation derived from the Latin* explanare*: Flatten, make level or plane (Harper-Collins Dictionary of Philosophy, 1992). 3) inter-packet coordination can involve structures residing outside packet network (i.e., cortico-thalamo–cortical connections, vs. cortico-cortical connections).*

#### **Figure 14.**

*The experience of mental effort. "Call the forgotten thing Z, the first facts with which we felt it was related to a, b, and c, and the details finally operative in calling it up 1, m, and n. The activity in Z will at first be a mere tension; but as the activities in a, b, and c little by little irradiate into l, m, and n … their combined irradiations upon Z succeed in helping the tension there to overcome the resistance, and in rousing Z to full activity. Through hovering of the attention in the neighborhood of the desired object, the accumulation of associates becomes so great that the combined tensions of their neural processes break through the bar, and the nervous wave pours into the tract, which has so long been awaiting its advent" ([56] p. 586).*

excitatory-inhibitory connections between the packets serving to constrain vector rotation. **Figure 15** illustrates these notions.

4.Mental modeling entered the stage (i.e., Sapience emerged) when mental processes became decoupled from the motor-sensory feedback. The hypothesis is that neuronal machinery of sensory-motor coordination richly developed in the protohuman was adopted for the task of mental coordination not accompanied by any overt activities [28]. As a result, neuronal mechanisms could retain a rich repertoire of coordination capabilities but became unencumbered by the spatio-temporal constraints facing sensory-motor acts (e.g., when raising a hand, one cannot skip over intermediate positions or exceed the range and speed limits afforded by the muscular-skeletal system. By contrast, envisioning the same act does not face such restrictions).

Decoupling from motor-sensory feedback created a gateway into mental universe populated by products of composition (imagination). To yield adaptive benefits, regulatory mechanisms were needed that would curtail superfluous compositions and facilitate those that could be mapped back onto and benefit overt behavior (i.e., allow predictions). Understanding is such a mechanism: although being rooted in sensory-motor coordination, understanding allows predictions unrestricted by spatio-temporal limitations of sensory-motor processes

#### **Figure 15.**

*Understanding chess positions. White knight can move to 8 squares, thinking of possible moves involves consecutive activation of one place neuron and inhibiting the other seven in the knight packet. Place neuron responding to square a in the white pawn packet inhibits the corresponding neuron in the knight packet. As a result, the idea of moving knight to square a does not come to mind. Place neuron responding to square b in the knight packet excites place neuron b in the pawn packet, and vice versa. As a result, the idea of taking the black pawn by either the white pawn or the white knight presents itself prominently (one 'sees' the opportunity).*

#### **Figure 16.**

*Functional architecture underlying active inference. The architecture comprises 6 levels, from subcellular to model networks. Subcellular networks at the bottom coordinate movement of mitochondria and substances across cell populations and inside cells. The model network on top comprises a multitude of mental models spreading across different tasks and domains. Interactions between levels are two directional: Intra-level processes form groups of elements that are treated as (composite) elements in the next level above; in turn, upper level-processes influence conditions and coordinate groupings in the level below. The packet network plays a pivotal role in the architecture, bridging levels shared by all species and those that are unique to the humans and become operational gradually in the course of an individual's cognitive development.*

or the speed of neuronal signaling. At the same time, mental models are subject to constraints of a different kind, including the explainability requirement and, crucially, limitations imposed by processes (reentrant mapping) that are inherent in the coordination mechanisms and allow eliminating superfluous degrees of freedom in the model constituents. **Figure 16** summarizes assumptions and suggestions in this part, presenting a sketch of functional hierarchy underlying active inference.

**93**

*Brain Functional Architecture and Human Understanding*

on them in assessing assess the validity of their conclusions.

ideas and might help assessing their biological plausibility.

tion references findings supporting key notions in this proposal.

and also at the level of brain networks" ([63] p. 252).

*4.3.2 Optimizing deployment of neuronal resources*

**4.3 Assessing plausibility**

*4.3.1 Tuning neuronal resources*

Emergence of packets underlies perception, i.e., extraction of quasi-stable, bounded feature groupings (objects) from the sensory stream (e.g., one can discern and subsequently recognize different chess pieces). The relational level is split in two – behavioral and relational proper. In the former, different behavior patterns are attributed to the objects (e.g., admissible moves are defined for knight, as in **Figure 15**). In the latter, inter-object relations get decoupled from the objects' sensory contents (e.g., coordinations in **Figure 15** make no account of the shape, color, weight, etc. of the participating pieces). Finally, operations in the model network support mental experiments (gedanken experiments) – a form of active inference most distant from the control of motorsensory feedback. Mental experiments can entail physical experiments but do not rely

Ideas and suggestions in this section do not answer questions **a** and **b** posed at the end of Section 2 but, arguably, indicate directions for further inquiry. Question **c** will be addressed briefly in the discussion. The ideas are speculative, the next section references findings and theories in the literature that seem to agree with the

A thumbnail summary of the preceding two sections: Cognitive processes yield adaptive behavior via two regulatory loops: the first loop optimizes (coordinates) deployment of sensory-motor resources while the second loop coordinates deployment of neuronal resources. The first loop produces associative networks that give rise to packet networks, the second loop combines packets into nested coordinated structures (mental models). The second loop was decoupled from the motor-sensory feedback, which created an opportunity for constructing unlimited multitudes of mental models. Realization of that opportunity was predicated on satisfying two constraints: a) using a limited number of neurons and b) maintaining energy consumption below some physiologically attainable thresholds. It can be shown that mechanisms of packets and packet coordination are deployment heuristics serving to satisfy the constraints [11, 12]. Packet coordination underlies understanding, which is a form of active inference unique to Sapience. The remainder of this sec-

Dynamic allocation of neuronal resources implies that neurons have a degree of plasticity, i.e. their receptive fields (RF) can be changed by both the stimuli and, crucially, brain systems that regulate allocation. A body of findings in [57–62] provide ample evidence of such plasticity, including stimulus-driven adaptive plasticity, rapid attention-driven plasticity, and consolidated learning-induced plasticity. Rapid attention-driven plasticity manifests in attentional modulation of neuronal processes and underlies the ability of the brain to make coordinated changes in stimuli-driven and self-directed neuronal activities as the context and task demands change. "These transformations occur at the level of synapses, single-neuron RFs,

The idea to characterize cognitive processes as resource optimization has been explored repeatedly in several forms, as optimization of energetic resources [64], optimization of computing resources [65], optimization of cognitive resources [66]. The present theory characterizes cognition as deployment of neuronal resources optimized

*DOI: http://dx.doi.org/10.5772/intechopen.95594*

Emergence of packets underlies perception, i.e., extraction of quasi-stable, bounded feature groupings (objects) from the sensory stream (e.g., one can discern and subsequently recognize different chess pieces). The relational level is split in two – behavioral and relational proper. In the former, different behavior patterns are attributed to the objects (e.g., admissible moves are defined for knight, as in **Figure 15**). In the latter, inter-object relations get decoupled from the objects' sensory contents (e.g., coordinations in **Figure 15** make no account of the shape, color, weight, etc. of the participating pieces). Finally, operations in the model network support mental experiments (gedanken experiments) – a form of active inference most distant from the control of motorsensory feedback. Mental experiments can entail physical experiments but do not rely on them in assessing assess the validity of their conclusions.

Ideas and suggestions in this section do not answer questions **a** and **b** posed at the end of Section 2 but, arguably, indicate directions for further inquiry. Question **c** will be addressed briefly in the discussion. The ideas are speculative, the next section references findings and theories in the literature that seem to agree with the ideas and might help assessing their biological plausibility.

## **4.3 Assessing plausibility**

*Connectivity and Functional Specialization in the Brain*

**92**

inference.

**Figure 16.**

**Figure 15.**

or the speed of neuronal signaling. At the same time, mental models are subject to constraints of a different kind, including the explainability requirement and, crucially, limitations imposed by processes (reentrant mapping) that are inherent in the coordination mechanisms and allow eliminating superfluous degrees of freedom in the model constituents. **Figure 16** summarizes assumptions and suggestions in this part, presenting a sketch of functional hierarchy underlying active

*Functional architecture underlying active inference. The architecture comprises 6 levels, from subcellular to model networks. Subcellular networks at the bottom coordinate movement of mitochondria and substances across cell populations and inside cells. The model network on top comprises a multitude of mental models spreading across different tasks and domains. Interactions between levels are two directional: Intra-level processes form groups of elements that are treated as (composite) elements in the next level above; in turn, upper level-processes influence conditions and coordinate groupings in the level below. The packet network plays a pivotal role in the architecture, bridging levels shared by all species and those that are unique to the humans and become operational gradually in the course of an individual's cognitive development.*

*Understanding chess positions. White knight can move to 8 squares, thinking of possible moves involves consecutive activation of one place neuron and inhibiting the other seven in the knight packet. Place neuron responding to square a in the white pawn packet inhibits the corresponding neuron in the knight packet. As a result, the idea of moving knight to square a does not come to mind. Place neuron responding to square b in the knight packet excites place neuron b in the pawn packet, and vice versa. As a result, the idea of taking the black pawn by either the white pawn or the white knight presents itself prominently (one 'sees' the opportunity).*

A thumbnail summary of the preceding two sections: Cognitive processes yield adaptive behavior via two regulatory loops: the first loop optimizes (coordinates) deployment of sensory-motor resources while the second loop coordinates deployment of neuronal resources. The first loop produces associative networks that give rise to packet networks, the second loop combines packets into nested coordinated structures (mental models). The second loop was decoupled from the motor-sensory feedback, which created an opportunity for constructing unlimited multitudes of mental models. Realization of that opportunity was predicated on satisfying two constraints: a) using a limited number of neurons and b) maintaining energy consumption below some physiologically attainable thresholds. It can be shown that mechanisms of packets and packet coordination are deployment heuristics serving to satisfy the constraints [11, 12]. Packet coordination underlies understanding, which is a form of active inference unique to Sapience. The remainder of this section references findings supporting key notions in this proposal.

#### *4.3.1 Tuning neuronal resources*

Dynamic allocation of neuronal resources implies that neurons have a degree of plasticity, i.e. their receptive fields (RF) can be changed by both the stimuli and, crucially, brain systems that regulate allocation. A body of findings in [57–62] provide ample evidence of such plasticity, including stimulus-driven adaptive plasticity, rapid attention-driven plasticity, and consolidated learning-induced plasticity. Rapid attention-driven plasticity manifests in attentional modulation of neuronal processes and underlies the ability of the brain to make coordinated changes in stimuli-driven and self-directed neuronal activities as the context and task demands change. "These transformations occur at the level of synapses, single-neuron RFs, and also at the level of brain networks" ([63] p. 252).

#### *4.3.2 Optimizing deployment of neuronal resources*

The idea to characterize cognitive processes as resource optimization has been explored repeatedly in several forms, as optimization of energetic resources [64], optimization of computing resources [65], optimization of cognitive resources [66]. The present theory characterizes cognition as deployment of neuronal resources optimized for energy efficiency, under an exceedingly simple model ("neurons fire at stimuli"): successful allocation of neurons to streaming stimuli procures energy deposits from the stimuli and incurs energy costs (recruiting, firing, maintaining neurons), neuronal system seeks to maximize the former while minimizing the latter [12]. It can be shown that, under this model, elements of functional architecture in **Figure 16** represent heuristics delivering progressively improving energy inflows while reducing energy costs (optimal maneuvering of neuronal resources to maximize gains and minimize losses). Other major phenomena can be mapped straightforwardly onto the model, e.g. in the context of resource optimization, the short term memory/long term memory partitioning turns out to be a powerful heuristic involving breaking large optimization problems into successions of small ones thus cutting down the amount of computation while keeping the outcome in the vicinity of global optimum. Optimal allocation strategies include prediction and anticipatory recruitment (active inference), combining those with cost minimization enabled expansion and diversification of inference domains. Dynamic resource optimization requires unencumbered access to all resources in the resource pool and flexible switching between resource groupings. These notions resonate with proposals in the literature, some examples follow.

A model in [67] postulates a global workspace composed of distributed and heavily interconnected neurons, and a set of specialized modules conducting perceptual, motor, evaluative, and attentional operations. Workspace (regulatory) neurons are mobilized in effortful tasks and selectively mobilize or suppress, through descending connections, the contribution of specific processor neurons. When workspace neurons become spontaneously co- activated, they form spatiotemporal patterns that are subject to modulation by vigilance signals.

The idea of cost-reward tradeoffs is consistent with the findings in [68]. This study examined neuronal substrate responsible for balancing expected performance rewards and their cognitive costs. Single-unit recordings in monkeys provided evidence that neurons in Medial Frontal Cortex (MFC) encode associations between action sets and their rewarding values and are involved in the cost- reward tradeoffs. MFC evaluates the costs incurred in executing cognitively demanding tasks and the expected gains, and recruits control resources in the Lateral Prefrontal Cortex (LPC) as necessary for compensating performance costs. MFC responses also reflect intrinsic MFC processes inhibiting inappropriate behaviors and energizing the LPC resources involved in selecting alternative behaviors according to the rewards and penalties at stake. The ideas concerning the cost-reward tradeoffs are consistent with those in [69].

The overall notion of dynamically optimized recruitment of neuronal resources is consistent with findings in [70] associating competent performance across multiple domains ("general intelligence") with selective recruitment of lateral frontal cortex in one or both hemispheres. These same frontal regions were found to be recruited by a broad range of cognitive demands, thus suggesting that "general intelligence" derives from flexibly switching recruitment between different neuronal groups. Another facet of neuronal processes implicit in the idea of neuronal resource optimization is "neuronal reuse", i.e. engaging the same circuitry for different behavioral purposes [71]. Combining quasi-stable neuronal packets without changing the packets or the underlying mosaic of associative links is a form of reuse. Improving energy efficiency can be a factor in the optimization of cerebral cortex layout and physical embedding of processing networks in the brain volume [72]): minimization of total connection length [73] reduces energy costs of signal propagation.

#### *4.3.3 Improving energy efficiency*

Neuronal processes consume significant amount of energy, consumption increases with activity which demands local and global changes in metabolic rates

**95**

*Brain Functional Architecture and Human Understanding*

and blood flow. Mechanisms of efficiency and energy transduction in the brain have been investigated in numerous studies [74–78]. Energy is produced through oxygen consumption mediated by the mitochondrial respiratory chain generating the high-energy phosphorous metabolite (adenosine triphosphate, or ATP). The carbon source that supports the oxidative metabolism is predominantly glucose. About 20% of the total oxygen consumption in the body takes place in the brain. A detailed account of energy consumption was obtained in a recent study utilizing 31P-MRS in vivo imaging of the human brain [79]. It was determined that approximately 5.7 kg of ATP molecules is produced and utilized by the cortical gray and white matter in a day, which is equivalent to the complete oxidative combustion of 56 g glucose per day and is almost five times the total weight of the gray and white matter (≈1.2 kg). The energy expenditure of a single cortical neuron is 4.7 billion ATPs per second (compared than 3.3 billion ATPs/neuron/sec estimated for the rat brain). Approximately 67–75% of the total energy expenditures is used for neurotransmitter signaling and electrophysiological activities involved in sustaining

It has been long recognized that the high energetic cost of human brain function, which is 10 times higher than what would be expected from its weight alone, can only be maintained through efficient energy use [80, 81]. Accordingly, theories were advanced suggesting that brains evolved to be metabolically efficient [82–84] which implies that representations of events and actions should be sculpted to involve as few action potentials and active synapses as possible. For optimum efficiency, less than 4% of a population of cortical neurons should be activated to represent a new event. Neural mechanisms associated with attention restrict the volume of cortex in which activity is elevated [85]. The arrangements of neuronal systems are thought to allow maximum communication speed with minimal energy

Massive data was accumulated demonstrating reduction of metabolic costs in the organization of motor performance and regulation of movement economy [87–93]. As noted in [94], metabolic determinants of physical action organization might not be the same as those determining cognitive action organization. However, it stands to reason to assume that the principle of cost minimization applies in both domains. Analysis in [85] concludes that strategies directed at maximizing metabolic efficiency are indeed used by the brain. In particular, a) fine axon collaterals reduce the number of ions required to transmit an action potential, by reducing membrane area, b) the arrangement of neurons in maps reduces the distance the potentials must travel and c) sparse codes reduce the number of action potentials required to represent events. **Figure 17** indicates that suggestions in the present theory resonate

The idea that dynamically formed neuronal groupings (assemblies, ensembles)

are the basic functional units in neuronal processes was advanced by [95] and subsequently developed in the Theory of Neuronal Group Selection (TNGS) by Gerald Edelman [96–98] and explored in other studies. For example, [99] suggests that acquisition of motor skills involves development of motor primitives amenable to adaptive re-combination (arguably, motor primitives are rooted in the underlying neuronal assemblies), [100] conceptualizes mental synthesis as a synchronization of independent neuronal ensembles, etc. Hebb's idea received experimental support in a number of recent findings: studies in [86, 101] demonstrated existence of neuronal assemblies entering into different combinations as the tasks and conditions change. Assemblies observed in [86] comprise a few dozen neurons each and can

*DOI: http://dx.doi.org/10.5772/intechopen.95594*

neuronal functions [79].

expenditures [86].

with those formulated in [85] and other studies.

*4.3.4 Neuronal packets are building blocks in cognitive processes*

#### *Brain Functional Architecture and Human Understanding DOI: http://dx.doi.org/10.5772/intechopen.95594*

*Connectivity and Functional Specialization in the Brain*

for energy efficiency, under an exceedingly simple model ("neurons fire at stimuli"): successful allocation of neurons to streaming stimuli procures energy deposits from the stimuli and incurs energy costs (recruiting, firing, maintaining neurons), neuronal system seeks to maximize the former while minimizing the latter [12]. It can be shown that, under this model, elements of functional architecture in **Figure 16** represent heuristics delivering progressively improving energy inflows while reducing energy costs (optimal maneuvering of neuronal resources to maximize gains and minimize losses). Other major phenomena can be mapped straightforwardly onto the model, e.g. in the context of resource optimization, the short term memory/long term memory partitioning turns out to be a powerful heuristic involving breaking large optimization problems into successions of small ones thus cutting down the amount of computation while keeping the outcome in the vicinity of global optimum. Optimal allocation strategies include prediction and anticipatory recruitment (active inference), combining those with cost minimization enabled expansion and diversification of inference domains. Dynamic resource optimization requires unencumbered access to all resources in the resource pool and flexible switching between resource groupings. These notions resonate with proposals in the literature, some examples follow. A model in [67] postulates a global workspace composed of distributed and heavily interconnected neurons, and a set of specialized modules conducting perceptual, motor, evaluative, and attentional operations. Workspace (regulatory) neurons are mobilized in effortful tasks and selectively mobilize or suppress, through descending connections, the contribution of specific processor neurons. When workspace neurons become spontaneously co- activated, they form spatio-

temporal patterns that are subject to modulation by vigilance signals.

The idea of cost-reward tradeoffs is consistent with the findings in [68]. This study examined neuronal substrate responsible for balancing expected performance rewards and their cognitive costs. Single-unit recordings in monkeys provided evidence that neurons in Medial Frontal Cortex (MFC) encode associations between action sets and their rewarding values and are involved in the cost- reward tradeoffs. MFC evaluates the costs incurred in executing cognitively demanding tasks and the expected gains, and recruits control resources in the Lateral Prefrontal Cortex (LPC) as necessary for compensating performance costs. MFC responses also reflect intrinsic MFC processes inhibiting inappropriate behaviors and energizing the LPC resources involved in selecting alternative behaviors according to the rewards and penalties at stake. The ideas concerning the cost-reward tradeoffs are consistent with those in [69].

The overall notion of dynamically optimized recruitment of neuronal resources is consistent with findings in [70] associating competent performance across multiple domains ("general intelligence") with selective recruitment of lateral frontal cortex in one or both hemispheres. These same frontal regions were found to be recruited by a broad range of cognitive demands, thus suggesting that "general intelligence" derives from flexibly switching recruitment between different neuronal groups. Another facet of neuronal processes implicit in the idea of neuronal resource optimization is "neuronal reuse", i.e. engaging the same circuitry for different behavioral purposes [71]. Combining quasi-stable neuronal packets without changing the packets or the underlying mosaic of associative links is a form of reuse. Improving energy efficiency can be a factor in the optimization of cerebral cortex layout and physical embedding of processing networks in the brain volume [72]): minimization

of total connection length [73] reduces energy costs of signal propagation.

Neuronal processes consume significant amount of energy, consumption increases with activity which demands local and global changes in metabolic rates

**94**

*4.3.3 Improving energy efficiency*

and blood flow. Mechanisms of efficiency and energy transduction in the brain have been investigated in numerous studies [74–78]. Energy is produced through oxygen consumption mediated by the mitochondrial respiratory chain generating the high-energy phosphorous metabolite (adenosine triphosphate, or ATP). The carbon source that supports the oxidative metabolism is predominantly glucose. About 20% of the total oxygen consumption in the body takes place in the brain. A detailed account of energy consumption was obtained in a recent study utilizing 31P-MRS in vivo imaging of the human brain [79]. It was determined that approximately 5.7 kg of ATP molecules is produced and utilized by the cortical gray and white matter in a day, which is equivalent to the complete oxidative combustion of 56 g glucose per day and is almost five times the total weight of the gray and white matter (≈1.2 kg). The energy expenditure of a single cortical neuron is 4.7 billion ATPs per second (compared than 3.3 billion ATPs/neuron/sec estimated for the rat brain). Approximately 67–75% of the total energy expenditures is used for neurotransmitter signaling and electrophysiological activities involved in sustaining neuronal functions [79].

It has been long recognized that the high energetic cost of human brain function, which is 10 times higher than what would be expected from its weight alone, can only be maintained through efficient energy use [80, 81]. Accordingly, theories were advanced suggesting that brains evolved to be metabolically efficient [82–84] which implies that representations of events and actions should be sculpted to involve as few action potentials and active synapses as possible. For optimum efficiency, less than 4% of a population of cortical neurons should be activated to represent a new event. Neural mechanisms associated with attention restrict the volume of cortex in which activity is elevated [85]. The arrangements of neuronal systems are thought to allow maximum communication speed with minimal energy expenditures [86].

Massive data was accumulated demonstrating reduction of metabolic costs in the organization of motor performance and regulation of movement economy [87–93]. As noted in [94], metabolic determinants of physical action organization might not be the same as those determining cognitive action organization. However, it stands to reason to assume that the principle of cost minimization applies in both domains.

Analysis in [85] concludes that strategies directed at maximizing metabolic efficiency are indeed used by the brain. In particular, a) fine axon collaterals reduce the number of ions required to transmit an action potential, by reducing membrane area, b) the arrangement of neurons in maps reduces the distance the potentials must travel and c) sparse codes reduce the number of action potentials required to represent events. **Figure 17** indicates that suggestions in the present theory resonate with those formulated in [85] and other studies.

### *4.3.4 Neuronal packets are building blocks in cognitive processes*

The idea that dynamically formed neuronal groupings (assemblies, ensembles) are the basic functional units in neuronal processes was advanced by [95] and subsequently developed in the Theory of Neuronal Group Selection (TNGS) by Gerald Edelman [96–98] and explored in other studies. For example, [99] suggests that acquisition of motor skills involves development of motor primitives amenable to adaptive re-combination (arguably, motor primitives are rooted in the underlying neuronal assemblies), [100] conceptualizes mental synthesis as a synchronization of independent neuronal ensembles, etc. Hebb's idea received experimental support in a number of recent findings: studies in [86, 101] demonstrated existence of neuronal assemblies entering into different combinations as the tasks and conditions change. Assemblies observed in [86] comprise a few dozen neurons each and can

#### **Figure 17.**

*Connected associative network allows unrestricted signal propagation, i.e., excitation of any neuron can ignite excitation spreading that will, eventually, engulf the entire network. Formation of packets and operations on them minimize spreading, confining excitation to the smallest subset of neurons producing the largest expected energy gain. Dynamic resource optimization boils down to suppressing wasteful firing and facilitating beneficial firing, i.e. yielding maximum prediction accuracy and response composition optimal under the prediction. In that sense, resource optimization is an engine of active inference.*

be interlaced within the same volume. It was suggested that "elementary neuronal groups are prescribed Lego-like building blocks of perception and that acquired memory relies on combining these elementary assemblies into higher-order constructs" [86]. Both studies suggest that their findings reveal a synaptic organizing principle (i.e., grouping) that is common across animals.

An important elaboration of the notion of assembly received in the idea of *synergistic structural units* formulated in [102–106]. Synergistic structural units can be combined into task- specific groupings and, crucially, are amenable to "nonindividualized control", that is, their constituent elements can be controlled by a few task-related variables (goals) [103].

The notion of 'neuronal packets' builds on the idea of Hebbian assembly and is consistent with the finding and suggestion referenced above. However, the notion offers two crucial extensions to the idea, as follows: a) neuronal packets form as a result of phase transition in associative networks causing some subnets to fold into cohesive units (packets) and b) folding establishes energy barriers at the packet boundaries. Stated differently, boundary energy barriers implement Markov blankets separating packet internals from the surrounding network [30]. More precisely, the height of energy barriers equals free energy per unit of surface area (surface tension) determined by the total membrane surface in the packet's boundary neurons (i.e., packet's Markov blanket, see **Figure 9**). Analysis in [85] identified reduction of membrane areas in individual neurons as a factor contributing into brain's metabolic efficiency. In a similar way, thermodynamically-driven tendency to minimize packet surface areas [27] contributes to the metabolic efficiency of neuronal processes (see **Figure 17**).

#### *4.3.5 Coordinated rotation of packet vectors*

Phase transitions transform groups of associated neurons into cohesive functional units amenable to synergistic control and re-combination with other units (reuse). The present theory defines coordinated rotation of packet vectors as a form of synergistic control, extending control mechanism described in [51–53, 107–111] from controlling overt movements to controlling mental 'movements' (i.e., packet vector rotation and coordination, see **Figure 15**). This generalization is consistent

**97**

sequences [117].

**Figure 18.**

*propagation (adopted from [95]).*

vector rotation and reentrant mapping.

*4.3.6 Reentrant mapping*

*Brain Functional Architecture and Human Understanding*

trajectories within the assembly, as shown in **Figure 18**.

with the original concept of neuronal assemblies in [95] envisioning the possibility of the assemblies producing different responses constituted by different excitation

*Neuronal assemblies were conceptualized as complex structures affording different trajectories for excitation* 

Cortical coordination and dynamics have been analyzed in [118–121] concluding that "the formation of neural context through the coordinated mutual constraint of multiple interacting cortical areas, is considered as a guiding principle underlying all cognitive functions" ([120] p. 140). The present theory agrees with that conclusion and suggests neuronal mechanisms instantiating the idea. In particular, the theory defines mental models as tightly coordinated gestalts, or *structural units* where changes in one component cause reciprocal changes in the other ones (e.g. when one hand is used to lift a heavy object from a tray supported by the other hand, increasing effort in one hand is concomitant with relaxation in the other one – hands form a structural unit (Gelfand et al). The same coordination mechanism underlies operation of mental models, e. g, in a catapult model, increasing distance to the target entails the realization that projectile need to be shifted in the opposite direction). The modeling mechanisms includes coordinated

The hypothesis that reentrant signaling serves as a general mechanism to facilitate the coordination of neuronal firing in anatomically and functionally segregated cortical areas and in the thalamus is one of the main tenets in the Theory of Neuronal Groups Selection (TNGS) [122–124]. According to TNGS, neurons

The notion of packet vector trajectories appears to be consistent with the findings in [112] demonstrating that memorization involves formation of specific sequences of spike bursts in the cortex that are replayed during retrieval. The function of coordination (neurons *Zk* in **Figure 15**) can be carried out by components of basal ganglia, thalamus and other structures. In particular, [113] suggests that basal ganglia chunks the representations of motor and cognitive action sequences so that they can be implemented as performance units. Studies in [114] uncovered activities in basal ganglia circuits that encoded sequences as single actions. Besides start/stop signaling and sequence parsing, these neurons displayed inhibited or sustained activity throughout the execution of the sequences. This sustained activity co-varied with the rate of execution of individual sequence elements, consistent with motor concatenation. Direct and indirect pathways of basal ganglia were concomitantly active during sequence initiation, but behaved differently during performance. Thalamic relays also play a critical role in coordination [115, 116]. The cerebellum is also involved in the detection and generation of

*DOI: http://dx.doi.org/10.5772/intechopen.95594*

*Brain Functional Architecture and Human Understanding DOI: http://dx.doi.org/10.5772/intechopen.95594*

#### **Figure 18.**

*Connectivity and Functional Specialization in the Brain*

be interlaced within the same volume. It was suggested that "elementary neuronal groups are prescribed Lego-like building blocks of perception and that acquired memory relies on combining these elementary assemblies into higher-order constructs" [86]. Both studies suggest that their findings reveal a synaptic organizing

*Connected associative network allows unrestricted signal propagation, i.e., excitation of any neuron can ignite excitation spreading that will, eventually, engulf the entire network. Formation of packets and operations on them minimize spreading, confining excitation to the smallest subset of neurons producing the largest expected energy gain. Dynamic resource optimization boils down to suppressing wasteful firing and facilitating beneficial firing, i.e. yielding maximum prediction accuracy and response composition optimal under the* 

An important elaboration of the notion of assembly received in the idea of *synergistic structural units* formulated in [102–106]. Synergistic structural units can be combined into task- specific groupings and, crucially, are amenable to "nonindividualized control", that is, their constituent elements can be controlled by a few

The notion of 'neuronal packets' builds on the idea of Hebbian assembly and is consistent with the finding and suggestion referenced above. However, the notion offers two crucial extensions to the idea, as follows: a) neuronal packets form as a result of phase transition in associative networks causing some subnets to fold into cohesive units (packets) and b) folding establishes energy barriers at the packet boundaries. Stated differently, boundary energy barriers implement Markov blankets separating packet internals from the surrounding network [30]. More precisely, the height of energy barriers equals free energy per unit of surface area (surface tension) determined by the total membrane surface in the packet's boundary neurons (i.e., packet's Markov blanket, see **Figure 9**). Analysis in [85] identified reduction of membrane areas in individual neurons as a factor contributing into brain's metabolic efficiency. In a similar way, thermodynamically-driven tendency to minimize packet surface areas [27] contributes to the metabolic efficiency of

Phase transitions transform groups of associated neurons into cohesive functional units amenable to synergistic control and re-combination with other units (reuse). The present theory defines coordinated rotation of packet vectors as a form of synergistic control, extending control mechanism described in [51–53, 107–111] from controlling overt movements to controlling mental 'movements' (i.e., packet vector rotation and coordination, see **Figure 15**). This generalization is consistent

principle (i.e., grouping) that is common across animals.

*prediction. In that sense, resource optimization is an engine of active inference.*

task-related variables (goals) [103].

**Figure 17.**

neuronal processes (see **Figure 17**).

*4.3.5 Coordinated rotation of packet vectors*

**96**

*Neuronal assemblies were conceptualized as complex structures affording different trajectories for excitation propagation (adopted from [95]).*

with the original concept of neuronal assemblies in [95] envisioning the possibility of the assemblies producing different responses constituted by different excitation trajectories within the assembly, as shown in **Figure 18**.

The notion of packet vector trajectories appears to be consistent with the findings in [112] demonstrating that memorization involves formation of specific sequences of spike bursts in the cortex that are replayed during retrieval. The function of coordination (neurons *Zk* in **Figure 15**) can be carried out by components of basal ganglia, thalamus and other structures. In particular, [113] suggests that basal ganglia chunks the representations of motor and cognitive action sequences so that they can be implemented as performance units. Studies in [114] uncovered activities in basal ganglia circuits that encoded sequences as single actions. Besides start/stop signaling and sequence parsing, these neurons displayed inhibited or sustained activity throughout the execution of the sequences. This sustained activity co-varied with the rate of execution of individual sequence elements, consistent with motor concatenation. Direct and indirect pathways of basal ganglia were concomitantly active during sequence initiation, but behaved differently during performance. Thalamic relays also play a critical role in coordination [115, 116]. The cerebellum is also involved in the detection and generation of sequences [117].

Cortical coordination and dynamics have been analyzed in [118–121] concluding that "the formation of neural context through the coordinated mutual constraint of multiple interacting cortical areas, is considered as a guiding principle underlying all cognitive functions" ([120] p. 140). The present theory agrees with that conclusion and suggests neuronal mechanisms instantiating the idea. In particular, the theory defines mental models as tightly coordinated gestalts, or *structural units* where changes in one component cause reciprocal changes in the other ones (e.g. when one hand is used to lift a heavy object from a tray supported by the other hand, increasing effort in one hand is concomitant with relaxation in the other one – hands form a structural unit (Gelfand et al). The same coordination mechanism underlies operation of mental models, e. g, in a catapult model, increasing distance to the target entails the realization that projectile need to be shifted in the opposite direction). The modeling mechanisms includes coordinated vector rotation and reentrant mapping.

#### *4.3.6 Reentrant mapping*

The hypothesis that reentrant signaling serves as a general mechanism to facilitate the coordination of neuronal firing in anatomically and functionally segregated cortical areas and in the thalamus is one of the main tenets in the Theory of Neuronal Groups Selection (TNGS) [122–124]. According to TNGS, neurons

belonging to different cortical areas are reciprocally interconnected by reentrant networks of excitatory axons, and each cortical area is also reentrantly interconnected by large numbers of axons to one or more nuclei of the thalamus. These thalamocortical and cortico-thalamic reentrant connections modulate brain arousal and help determining which of the patterns of environmental signals arriving in the thalamus from the environment will be relayed on to the cortex. They also participate in the execution of timed, sequential, or willed processes, such as manipulating mental constructs, or issuing segmented motor commands [124]. The present theory is consistent with TNGS principles, making reentrant mapping (or bidirectional coupling [113]) integral to the mechanisms of modeling and understanding (see **Figure 15**).

#### *4.3.7 Energy landscapes – a missing link in cognitive neuroscience*

It has been long recognized that the concept on neuronal assembly leaves the issues of stability and borders undetermined (how does the brain 'know' where one assembly ends and another begins, how does a neuron 'know' to which assembly it belongs, what keeps neurons in an assembly together, etc.)? In the original conceptualization [95], waves of excitations develop and reverberate inside assemblies - this notion indicates intuition of assembly borders but that intuition was not made explicit. The original conceptualization in [95] entailed a possibility that activity in any assembly will spread to other assemblies and ultimately to the entire cortex or even the total brain, resulting in pathological overactivity, as in seizures. To cope with the problem, the idea of a "threshold control mechanism" was introduced [125] with the subsequent elaborations placing the mechanism in the basal ganglia or the hippocampus. The idea was that a cell assembly "holds" at a threshold θ when at that threshold all the neurons of the assembly, once excited, stay active due to their reciprocal excitatory connections. Manipulation of the thresholds was envisioned as follows (compare to **Figure 13**(1)).

*"A periodic operation (colloquially called the "pump of thoughts") may involve the following steps. Given a certain input I, the threshold is lowered so that the set of active neurons FI will go over into a larger set F'I. This will encourage the ignition of cell assemblies. As the threshold is again raised, activity is smothered and only the most strongly connected cell assembly will survive. A new cycle beginning again with a lowered threshold will bring in new cell assemblies. They may include an even more strongly connected cell assembly, which will be the next one to survive when the threshold is raised. The evolution will be in the direction of the most strongly connected cell assemblies…. One may express this by saying that the system hunts for an interpretation of the input, or that it 'thinks" ([125] p.177).*

Independently from the proposal in [125], the idea of threshold regulation was advanced in a theory of movement coordination (λ theory) in [126–128] According to λ theory theory, coordination of motor actions involves centrally controlled resetting of the threshold positions of body segments. Deviations from the threshold positions (e.g., restive muscle length) trigger resistive forces, detection of differences between the centrally set threshold positions and the sensory-signaled actual positions cause activation of neuromuscular elements seeking to diminish the difference. The crucial assumption is that thresholds are changed by descending fibers that influence membrane potentials of motoneurons in motor cortices, either directly or via interneurons [126].

Arguably, theories of threshold regulation [125, 126] are motivated by intuition similar to that expressed in **Figures 13**-**15**. In the present theory, boundary barriers

**99**

*Brain Functional Architecture and Human Understanding*

are an intrinsic property of neuronal assemblies (packets), regulation of barrier height involves changes in membrane potential in neurons residing in the packet's MB. Boundary energy barriers make assemblies distinct, quasi-stable and immersed in energy landscapes. The landscape curtails activation spreading, by imposing

More precisely, the present theory postulates that boundary barriers establish energy landscapes across packet networks [10, 12]. Accordingly, formation of packets can be viewed as a form of folding, analogous to the folding of proteins and other complex molecular structures [129–131]. As in proteins, the folding of packets is a spontaneous process obtaining stable (equilibrium) configurations of minimal free energy [27]. Stability is maintained within some ranges of temperature varia-

> σ

become absorbed into the surrounding packets). Within the multidimensional energy surface, packets' Markov Blankets and the corresponding cutsets (links connecting MB to the surrounding packets) form attraction basins in the neighborhood of local minima, connected by saddle points. As a result, attentive navigation of the landscape involves energy-demanding (effortful) basin-to-basin transitions. Deformations in the energy landscape determine changes in the accessibility of neuronal packets. Presumably, transitions are controlled by frontal /prefrontal

A number of experimental results appear to agree with the proposal. Findings in [132] demonstrating fast transitions between separated states of cortical activity involving distinct neuronal groups appear to agree with above proposal. Findings in [133] indicate that thalamic cells respond selectively to complex percepts and concepts conferred on them by the cortical assemblies in whose activation they participate. The cortico-thalamo-cortical pathways provide connections between different cortical loci which have higher reliability than the direct cortico-cortical routes, and play crucial role in orchestrating activation of those assemblies). Important findings in [134] demonstrated that brain network are structured in a manner optimized for network control, which includes increased controllability and reduced synchronizability (controllability characterizes the ease of switching from one dynamical state to another, traversing energy landscape (see Figure 130; synchronizability characterizes the ability for regions in the network to support the

The idea of energy landscapes in brain systems remained purely speculative until the recent pioneering studies in [135–137] applied modern analytic and modeling techniques (e.g. network disconnectivity analysis) to fMRI data, seeking to define energy landscapes in Default Mode Network (DMN) and Fronto Parietal Network (FPN). It was determined that DMN energy landscape consisted of two groups of low-energy local minima that are separated by a relatively high energy barrier. Within each group, the activity patterns of the local minima were similar, and different minima were connected by relatively low energy barriers. In the FPN, all dominant local minima were separated by relatively low energy barriers such that they formed a single coarse-grained global minimum. The height of energy barriers separating local minima influences the rate of inter-state transitions. Accordingly, transitions in DNM occur at a low rate while transitions between local minima in FPN occur more easily. The notion that brain operates at the edge of instability and transits between low energy states has been explored in multiple studies [50, 138]). It appears that the notion of brain energy landscapes was introduced in [10, 12], and experimental

mapping of energy landscapes was attempted for the first time in [136].

To summarize, the folding of subnets in associative networks forms packets separated by Markov Blankets from the rest of the network. Packet Markov Blankets are constituted by boundary energy barriers that make packets distinct, quasi-stable

(*T T d T dT* ) ≈ ( )/ , *Em* → 0, the constituent neurons

*DOI: http://dx.doi.org/10.5772/intechopen.95594*

energy costs on inter-assembly transitions.

σ

tion (packets dissolve when

networks and thalamic structures.

same temporal dynamical patterns).

*Connectivity and Functional Specialization in the Brain*

*4.3.7 Energy landscapes – a missing link in cognitive neuroscience*

thresholds was envisioned as follows (compare to **Figure 13**(1)).

(see **Figure 15**).

belonging to different cortical areas are reciprocally interconnected by reentrant networks of excitatory axons, and each cortical area is also reentrantly interconnected by large numbers of axons to one or more nuclei of the thalamus. These thalamocortical and cortico-thalamic reentrant connections modulate brain arousal and help determining which of the patterns of environmental signals arriving in the thalamus from the environment will be relayed on to the cortex. They also participate in the execution of timed, sequential, or willed processes, such as manipulating mental constructs, or issuing segmented motor commands [124]. The present theory is consistent with TNGS principles, making reentrant mapping (or bidirectional coupling [113]) integral to the mechanisms of modeling and understanding

It has been long recognized that the concept on neuronal assembly leaves the issues of stability and borders undetermined (how does the brain 'know' where one assembly ends and another begins, how does a neuron 'know' to which assembly it belongs, what keeps neurons in an assembly together, etc.)? In the original conceptualization [95], waves of excitations develop and reverberate inside assemblies - this notion indicates intuition of assembly borders but that intuition was not made explicit. The original conceptualization in [95] entailed a possibility that activity in any assembly will spread to other assemblies and ultimately to the entire cortex or even the total brain, resulting in pathological overactivity, as in seizures. To cope with the problem, the idea of a "threshold control mechanism" was introduced [125] with the subsequent elaborations placing the mechanism in the basal ganglia or the hippocampus. The idea was that a cell assembly "holds" at a threshold θ when at that threshold all the neurons of the assembly, once excited, stay active due to their reciprocal excitatory connections. Manipulation of the

*"A periodic operation (colloquially called the "pump of thoughts") may involve the following steps. Given a certain input I, the threshold is lowered so that the set of active neurons FI will go over into a larger set F'I. This will encourage the ignition of cell assemblies. As the threshold is again raised, activity is smothered and only the most strongly connected cell assembly will survive. A new cycle beginning again with a lowered threshold will bring in new cell assemblies. They may include an even more strongly connected cell assembly, which will be the next one to survive when the threshold is raised. The evolution will be in the direction of the most strongly connected cell assemblies…. One may express this by saying that the system* 

*hunts for an interpretation of the input, or that it 'thinks" ([125] p.177).*

Independently from the proposal in [125], the idea of threshold regulation was advanced in a theory of movement coordination (λ theory) in [126–128] According to λ theory theory, coordination of motor actions involves centrally controlled resetting of the threshold positions of body segments. Deviations from the threshold positions (e.g., restive muscle length) trigger resistive forces, detection of differences between the centrally set threshold positions and the sensory-signaled actual positions cause activation of neuromuscular elements seeking to diminish the difference. The crucial assumption is that thresholds are changed by descending fibers that influence membrane potentials of motoneurons in motor cortices, either

Arguably, theories of threshold regulation [125, 126] are motivated by intuition similar to that expressed in **Figures 13**-**15**. In the present theory, boundary barriers

**98**

directly or via interneurons [126].

are an intrinsic property of neuronal assemblies (packets), regulation of barrier height involves changes in membrane potential in neurons residing in the packet's MB. Boundary energy barriers make assemblies distinct, quasi-stable and immersed in energy landscapes. The landscape curtails activation spreading, by imposing energy costs on inter-assembly transitions.

More precisely, the present theory postulates that boundary barriers establish energy landscapes across packet networks [10, 12]. Accordingly, formation of packets can be viewed as a form of folding, analogous to the folding of proteins and other complex molecular structures [129–131]. As in proteins, the folding of packets is a spontaneous process obtaining stable (equilibrium) configurations of minimal free energy [27]. Stability is maintained within some ranges of temperature variation (packets dissolve when σ σ (*T T d T dT* ) ≈ ( )/ , *Em* → 0, the constituent neurons become absorbed into the surrounding packets). Within the multidimensional energy surface, packets' Markov Blankets and the corresponding cutsets (links connecting MB to the surrounding packets) form attraction basins in the neighborhood of local minima, connected by saddle points. As a result, attentive navigation of the landscape involves energy-demanding (effortful) basin-to-basin transitions. Deformations in the energy landscape determine changes in the accessibility of neuronal packets. Presumably, transitions are controlled by frontal /prefrontal networks and thalamic structures.

A number of experimental results appear to agree with the proposal. Findings in [132] demonstrating fast transitions between separated states of cortical activity involving distinct neuronal groups appear to agree with above proposal. Findings in [133] indicate that thalamic cells respond selectively to complex percepts and concepts conferred on them by the cortical assemblies in whose activation they participate. The cortico-thalamo-cortical pathways provide connections between different cortical loci which have higher reliability than the direct cortico-cortical routes, and play crucial role in orchestrating activation of those assemblies). Important findings in [134] demonstrated that brain network are structured in a manner optimized for network control, which includes increased controllability and reduced synchronizability (controllability characterizes the ease of switching from one dynamical state to another, traversing energy landscape (see Figure 130; synchronizability characterizes the ability for regions in the network to support the same temporal dynamical patterns).

The idea of energy landscapes in brain systems remained purely speculative until the recent pioneering studies in [135–137] applied modern analytic and modeling techniques (e.g. network disconnectivity analysis) to fMRI data, seeking to define energy landscapes in Default Mode Network (DMN) and Fronto Parietal Network (FPN). It was determined that DMN energy landscape consisted of two groups of low-energy local minima that are separated by a relatively high energy barrier. Within each group, the activity patterns of the local minima were similar, and different minima were connected by relatively low energy barriers. In the FPN, all dominant local minima were separated by relatively low energy barriers such that they formed a single coarse-grained global minimum. The height of energy barriers separating local minima influences the rate of inter-state transitions. Accordingly, transitions in DNM occur at a low rate while transitions between local minima in FPN occur more easily. The notion that brain operates at the edge of instability and transits between low energy states has been explored in multiple studies [50, 138]). It appears that the notion of brain energy landscapes was introduced in [10, 12], and experimental mapping of energy landscapes was attempted for the first time in [136].

To summarize, the folding of subnets in associative networks forms packets separated by Markov Blankets from the rest of the network. Packet Markov Blankets are constituted by boundary energy barriers that make packets distinct, quasi-stable stable (i.e., amenable to modification but at substantial energy cost) and synergistic (i.e. amenable to control by a few variables and coordination with other packets). Boundary barriers establishes energy landscapes across packet networks and determine both kinematic (inter-packet transitions) and dynamic properties of neuronal organization.

#### *4.3.8 Accommodations*

It was suggested that lateral inhibition prevents neuronal assemblies from encroaching on each other while the tendency towards reducing surface tension in the packets favors their coalescence (minimizing the amount of free energy in the surface). Arguably, the interplay of the opposite tendencies drives 'accommodation', that is, spontaneous adjustments inside the neuronal systems following changes resulting from interactions with the environment [27].

The notions of assimilation, accommodation and cognitive equilibration were introduced in [18] denoting, correspondingly, integration of new information into the existing structures, re-organization of those structures, until a state of equilibrium is reached obtaining a sufficient degree of integration via a minimal amount of structural changes. According to the present theory, assimilation involves changes in the distribution of synaptic weights, that trigger waves of packet re-structuring propagating throughout the packet network (the accommodation). In this way, the requirement of spontaneous re-structuring is inherent in the notion of neuronal packets immersed in energy landscapes. On that view, the overall functional architecture of the cognitive systems was reduced to three modules: associative cortices, reticular formation controlling arousal level, and a frontal/prefrontal module controlling landscape navigation. Accommodation and assimilation are confined to packet networks [10].

Recent experimental findings and theoretical proposals [139] envision functional architecture comprising Default Mode Network (DMN) [140], Salience Network (SN) [141] and Task Control Network (TCN) [142], as follows. A DMN is a large network comprises hubs in medial prefrontal cortex, posterior cingulate/ precuneus and angular gyrus becomes active under conditions of wakeful rest, i.e. when person is not engaged in any task. The SN comprises a suite of brain regions whose cortical hubs are the anterior cingulate and ventral anterior insular cortices while the TCN (a cingulo-opercular task-control network) is anchored in the dorsal anterior insula and the frontal operculum. SN detects behaviorally relevant stimuli and recruits neural resources to orchestrate responses. For the latter, the SN engages the TCN (or Central Executive) whose functions include maintaining relevant task set or orchestrate switching to a new task set in response to shifts in the salience landscape.

Significantly, a comprehensive study in [143] compared functional networks in the brain during task performance (active brain) and at rest (resting brain), concluding that the full repertoire of functional networks utilized in active brain (Active Brain Networks, or ABN) remains continuously active in the resting brain (Resting State Networks, or RSN, including the "default mode network"). The study applied independent component analysis (ICA) and other modern techniques to two sets of fMRI functional imaging data: "active brain" data in the BrainMap data base collected from over 30,000 subjects, and resting brain data collected from 36 subjects. The ICA decomposition was conducted at two resolution levels, 20-component analysis and 70-component analysis, with the higher resolution analysis revealing subnetworks in the primary networks determined at the lower resolution level. It was found that primary networks split into subnetworks in both active and resting data in almost identical ways, maintaining greater functional

**101**

**Figure 19.**

*functional organization [147–149].*

*Brain Functional Architecture and Human Understanding*

(temporal) correlation between subnetworks within a primary network than across primary networks. Analysis in both levels produced converging results: close to 70% overlap in the composition of Active State Network and Resting State Networks. The analysis concludes with an admission: "Although we have shown that activation networks are mirrored in resting data, we must acknowledge that this does not begin to answer the question of why the brain's many regions continue to "function" (with large amplitude fluctuations) when the subject is at rest, and even when the

It appears that these findings are consistent with the proposal in [10] envision-

**Figure 19** suggests that DMN/SN/CEN interplay focuses on the engagement of prefrontal areas in coordination activities, i.e., formation of relations and operations on relational networks. Accordingly, it can be expected that prefrontal damages are likely to cause severe deficits in integration of relations. The order of operation in the DMN/SN/CEN system is, roughly, as follows: a) the Central Executive Network includes the agency of attention and controls attention focusing and other processes engaged in the performance of cognitive tasks, b) the Default Mode Network becomes active when the person remains awake but no tasks are pursued, c) the Salience Network administers switching between CEN

*Operations on networks underlying understanding capacity involve an interplay between default mode network and central executive network (CEN). Salience network coordinates switching between DMN and CEN [146]. This general architecture allows more detailed mapping onto anatomical structures in the brain underlying* 

ing waves of accommodating adjustments in packet networks. Moreover, the adjustment requirements are inherent in the notion of packet networks. In particular, the hypothesis is that variations in temperature and synaptic weight distribution across packet networks cause changes in the resting membrane potentials [144] in the MB neurons, thus creating potential gradients in the packet network causing adjustments in the energy landscape and re-distribution of neurons seeking packet configurations in the vicinity of global energy minima. Stated differently, energy landscape is "frustrated" [131] due to conflicting tendencies of lateral inhibition and lateral coalescence. Spontaneous re-organizations in packet networks to resolve frustration move the system in the direction of cognitive equilibrium. Possibly, neuronal avalanches are a form of such re-organization, playing a role in maintaining network stability and preventing runaway excitation [145]. **Figure 19** makes suggestions regarding the placement of packet networks in the tri-partite architec-

*DOI: http://dx.doi.org/10.5772/intechopen.95594*

subject is asleep and under anesthesia" [143].

ture [139, 146].

and DMN.

#### *Brain Functional Architecture and Human Understanding DOI: http://dx.doi.org/10.5772/intechopen.95594*

*Connectivity and Functional Specialization in the Brain*

resulting from interactions with the environment [27].

organization.

*4.3.8 Accommodations*

packet networks [10].

stable (i.e., amenable to modification but at substantial energy cost) and synergistic (i.e. amenable to control by a few variables and coordination with other packets). Boundary barriers establishes energy landscapes across packet networks and determine both kinematic (inter-packet transitions) and dynamic properties of neuronal

It was suggested that lateral inhibition prevents neuronal assemblies from encroaching on each other while the tendency towards reducing surface tension in the packets favors their coalescence (minimizing the amount of free energy in the surface). Arguably, the interplay of the opposite tendencies drives 'accommodation', that is, spontaneous adjustments inside the neuronal systems following changes

The notions of assimilation, accommodation and cognitive equilibration were introduced in [18] denoting, correspondingly, integration of new information into the existing structures, re-organization of those structures, until a state of equilibrium is reached obtaining a sufficient degree of integration via a minimal amount of structural changes. According to the present theory, assimilation involves changes in the distribution of synaptic weights, that trigger waves of packet re-structuring propagating throughout the packet network (the accommodation). In this way, the requirement of spontaneous re-structuring is inherent in the notion of neuronal packets immersed in energy landscapes. On that view, the overall functional architecture of the cognitive systems was reduced to three modules: associative cortices, reticular formation controlling arousal level, and a frontal/prefrontal module controlling landscape navigation. Accommodation and assimilation are confined to

Recent experimental findings and theoretical proposals [139] envision functional architecture comprising Default Mode Network (DMN) [140], Salience Network (SN) [141] and Task Control Network (TCN) [142], as follows. A DMN is a large network comprises hubs in medial prefrontal cortex, posterior cingulate/ precuneus and angular gyrus becomes active under conditions of wakeful rest, i.e. when person is not engaged in any task. The SN comprises a suite of brain regions whose cortical hubs are the anterior cingulate and ventral anterior insular cortices while the TCN (a cingulo-opercular task-control network) is anchored in the dorsal anterior insula and the frontal operculum. SN detects behaviorally relevant stimuli and recruits neural resources to orchestrate responses. For the latter, the SN engages the TCN (or Central Executive) whose functions include maintaining relevant task set or orchestrate switching to a new task set in response to shifts in the salience

Significantly, a comprehensive study in [143] compared functional networks in the brain during task performance (active brain) and at rest (resting brain), concluding that the full repertoire of functional networks utilized in active brain (Active Brain Networks, or ABN) remains continuously active in the resting brain (Resting State Networks, or RSN, including the "default mode network"). The study applied independent component analysis (ICA) and other modern techniques to two sets of fMRI functional imaging data: "active brain" data in the BrainMap data base collected from over 30,000 subjects, and resting brain data collected from 36 subjects. The ICA decomposition was conducted at two resolution levels, 20-component analysis and 70-component analysis, with the higher resolution analysis revealing subnetworks in the primary networks determined at the lower resolution level. It was found that primary networks split into subnetworks in both active and resting data in almost identical ways, maintaining greater functional

**100**

landscape.

(temporal) correlation between subnetworks within a primary network than across primary networks. Analysis in both levels produced converging results: close to 70% overlap in the composition of Active State Network and Resting State Networks. The analysis concludes with an admission: "Although we have shown that activation networks are mirrored in resting data, we must acknowledge that this does not begin to answer the question of why the brain's many regions continue to "function" (with large amplitude fluctuations) when the subject is at rest, and even when the subject is asleep and under anesthesia" [143].

It appears that these findings are consistent with the proposal in [10] envisioning waves of accommodating adjustments in packet networks. Moreover, the adjustment requirements are inherent in the notion of packet networks. In particular, the hypothesis is that variations in temperature and synaptic weight distribution across packet networks cause changes in the resting membrane potentials [144] in the MB neurons, thus creating potential gradients in the packet network causing adjustments in the energy landscape and re-distribution of neurons seeking packet configurations in the vicinity of global energy minima. Stated differently, energy landscape is "frustrated" [131] due to conflicting tendencies of lateral inhibition and lateral coalescence. Spontaneous re-organizations in packet networks to resolve frustration move the system in the direction of cognitive equilibrium. Possibly, neuronal avalanches are a form of such re-organization, playing a role in maintaining network stability and preventing runaway excitation [145]. **Figure 19** makes suggestions regarding the placement of packet networks in the tri-partite architecture [139, 146].

**Figure 19** suggests that DMN/SN/CEN interplay focuses on the engagement of prefrontal areas in coordination activities, i.e., formation of relations and operations on relational networks. Accordingly, it can be expected that prefrontal damages are likely to cause severe deficits in integration of relations. The order of operation in the DMN/SN/CEN system is, roughly, as follows: a) the Central Executive Network includes the agency of attention and controls attention focusing and other processes engaged in the performance of cognitive tasks, b) the Default Mode Network becomes active when the person remains awake but no tasks are pursued, c) the Salience Network administers switching between CEN and DMN.

#### **Figure 19.**

*Operations on networks underlying understanding capacity involve an interplay between default mode network and central executive network (CEN). Salience network coordinates switching between DMN and CEN [146]. This general architecture allows more detailed mapping onto anatomical structures in the brain underlying functional organization [147–149].*

### *4.3.9 Decoupling*

The present theory attributes emergence of human understanding to evolutionary developments causing decoupling of regulatory processes from the sensorymotor feedback loops [10, 28]. The idea is consistent with suggestions in [150] regarding evolutionary origins of human cognition. Analysis [150] focuses on the development of cerebral cortex, pointing at its vast expansion in the humans relative to other primates (the cerebral surface area is 120 cm2 in the macaque and 960 cm2 in the human) and disproportionate expansion of distributed association regions within the cortex. The hypothesis is that rapid expansion of the cortical mantle may have decoupled ("untethered") large portions of the cortex from sensory hierarchies and resulted in the development of networks that either control processes in the sensory networks or are engaged in parallel activities that are "detached from sensory perception and motor actions – what one might term 'internal mentation" [150].
