Guided Activation Theory of Prefrontal Function – Earl K. Miller & Jonathan D. Cohen

The quest to understand how the human brain orchestrates complex, goal-directed behavior in the face of an ever-changing and distracting environment represents one of the most enduring challenges in modern cognitive neuroscience. For decades, the prefrontal cortex was shrouded in mystery, frequently characterized by early clinical neurologists as the “silent cortex” because localized damage rarely yielded the striking sensory or gross motor deficits typical of primary cortical lesions. Instead, prefrontal damage manifested in subtler, yet catastrophic, breakdowns of volition, long-term planning, social comportment, and adaptive decision-making. As experimental psychology and cognitive science matured throughout the latter half of the twentieth century, conceptual constructs such as the “central executive” and the “supervisory attentional system” emerged to explain how the mind overrides automatic routines to pursue abstract intentions. Yet, these early psychological models suffered from a fundamental conceptual flaw: they frequently relied on an implicit homunculus—an ungrounded, intelligent agent inside the executive machinery that decides what to attend to, when to switch tasks, and how to govern action.

The turning point in resolving this homuncular dilemma arrived with the convergence of computational neuroscience, single-unit electrophysiology in non-human primates, and functional neuroimaging. In their landmark 2001 treatise published in the Annual Review of Neuroscience, Earl K. Miller and Jonathan D. Cohen formulated the Guided Activation Theory of Prefrontal Function. Rather than positing the prefrontal cortex as a supreme command center that directly generates motor acts or micro-manages every neural computation, Miller and Cohen conceptualized the prefrontal cortex as a source of continuous, top-down modulatory bias. In this framework, sensory and motor processing takes place within highly distributed, parallel pathways across the posterior and subcortical networks. When routine actions are triggered by strong, overlearned environmental stimuli, these posterior sensory-to-motor pathways operate semi-autonomously through hardwired associative weights. However, when behavioral contexts demand novel actions, the resolution of conflict, or the execution of counter-intuitive rules, prefrontal representations of goals and task contingencies inject top-down biasing signals that modulate the flow of neural activity along ambiguous pathways, systematically favoring behaviorally relevant computations over prepotent, automatic habits.

By conceptualizing cognitive control as an emergent property of top-down biasing within a competitive, distributed neural architecture, the Guided Activation Theory successfully bridged the chasm separating formal psychological theories of executive function from biophysically grounded neurobiology. Over the past two decades, this model has served as the foundational bedrock for contemporary research on cognitive flexibility, working memory, attention, and rule representation. It provides mechanistic explanations for human psychopathology, inspires architectures in artificial intelligence and deep reinforcement learning, and continues to evolve alongside advances in population-level neurodynamics, multi-regional electrophysiology, and high-dimensional neural geometry. This article provides a comprehensive and exhaustive exploration of the Guided Activation Theory, tracing its historical lineage, elucidating its core computational and biophysical mechanisms, examining its empirical validation across species, and evaluating its contemporary revisions at the cutting edge of modern systems neuroscience.

1. Historical Context and Theoretical Foundations of Prefrontal Cortex Research

1.1 Early Conceptualizations of the Prefrontal Cortex and Executive Function

The scientific journey toward understanding the prefrontal cortex (PFC) began in the mid-nineteenth century, marked by dramatic clinical observations that disrupted existing phrenological and radical localizationist doctrines. The classic case of Phineas Gage in 1848—a railroad foreman who survived the passage of an iron tamping rod through his left frontal lobe—demonstrated to the medical world that the frontal lobes were not essential for basic vegetative survival, language comprehension, or motor locomotion, but were indispensable for the regulation of personality, social convention, foresight, and impulse control. Gage transitioned from an industrious, mild-mannered professional into an erratic, irreverent, and profoundly undisciplined individual, incapable of executing planned endeavors.

Despite such dramatic clinical presentations, subsequent decades of neurosurgical and experimental investigation yielded perplexing findings. Mid-twentieth-century industrial applications of prefrontal lobotomies, pioneered by Egas Moniz and widely practiced by Walter Freeman, frequently left patients without overt perceptual deficits, intellectual decrements on standard IQ tests, or language aphasias. Consequently, the frontal lobes earned the moniker of the “silent cortex.” Neurophysiologists struggled to identify specific receptive fields or localized cognitive faculties within these vast anterior cortical expanses. Early experimental psychologists, working within rigid behaviorist paradigms, lacked the conceptual vocabulary to describe how internal states, abstract rules, and contextual knowledge could govern behavioral output independently of immediate sensory elicitation.

A major theoretical breakthrough emerged through the clinical-neuropsychological investigations of Alexander Luria. Operating in the mid-twentieth century, Luria proposed that the brain is organized into three primary functional units: the first unit regulates cortical tone and arousal (the brainstem reticular activating system); the second unit obtains, processes, and stores information from the external world (posterior sensory cortices); and the third unit is dedicated to programming, regulating, and verifying mental activity and behavior (the frontal lobes). Luria demonstrated that patients with substantial prefrontal pathology exhibited profound disturbances in complex programming: while their elemental motor components were structurally intact, they could not coordinate sequences of action according to verbal instructions, nor could they self-monitor errors or suppress automatic motor perseveration. Simultaneously, cognitive psychologists like Donald Broadbent and Donald Norman began formulating information-processing models that recognized the need for selective filters and supervisory attentional systems to gate the torrential cascade of sensory inputs confronting an organism.

1.2 The Supervisory Attentional System Framework of Norman and Shallice

Building directly upon Luria’s clinical observations and early information-processing paradigms, Donald Norman and Tim Shallice (1986) articulated the Supervisory Attentional System (SAS) framework, which became the immediate cognitive precursor to modern neurocomputational theories of executive control. Norman and Shallice sought to explain how the human cognitive architecture balances routine, automatic actions against deliberate, goal-directed interventions. They proposed a dual-tier processing architecture consisting of Contention Scheduling and the Supervisory Attentional System.

Contention scheduling was conceptualized as a decentralized, automatic network of action schemas—interconnected cognitive or motor subroutines refined through experience. When an individual engages in habitual behaviors, such as driving a familiar route or typing on a keyboard, perceptual cues activate corresponding schemas. These schemas compete with one another via mutual, lateral inhibition. The schema receiving the highest degree of bottom-up sensory reinforcement and environmental trigger support wins the competition, suppressing rival schemas and triggering motor execution without requiring conscious, effortful intervention. Contention scheduling is computationally efficient, rapid, and capable of coordinating multiple parallel actions, provided the environment conforms to established routines.

However, when an unexpected hazard arises, when novel environments demand non-routine actions, or when conflicting goals must be adjudicated, contention scheduling fails. Under these conditions, the Supervisory Attentional System intervenes. The SAS was defined as a higher-order, slow, capacity-limited controller capable of providing top-down modulation. Rather than directly executing actions, the SAS alters the activation levels of competing schemas within the contention scheduling network, artificially boosting goal-congruent subroutines or dampening prepotent, automatic responses. While the SAS model was brilliant in characterizing phenomena such as action slips, frontal dysexecutive syndrome, and voluntary task switching, it remained an abstract, box-and-arrow psychological construct. It did not define how individual neurons or physiological networks realize contention scheduling or supervisory interventions, leaving open the homunculus problem: what algorithmic principles decide how the SAS itself operates? Grounding this abstract model in biophysically plausible, cell-level neuroscience was the exact challenge taken up by Miller and Cohen.

1.3 The Working Memory Paradigm of Alan Baddeley and Patricia Goldman-Rakic

In parallel with the development of supervisory attentional models, the cognitive psychology of memory underwent an intellectual revolution through the formulation of the multi-component working memory model by Alan Baddeley and Graham Hitch (1974). Baddeley replaced the older concept of a passive, unitary short-term memory store with an active, dynamic workspace. The Baddeley architecture partitioned working memory into domain-specific “slave” storage systems—the phonological loop for auditory-verbal information and the visuospatial sketchpad for spatial arrays—coordinated and directed by a fractionated “central executive.” The central executive was assigned the responsibilities of binding information, focusing selective attention, switching between tasks, and coordinating mental transformations. However, much like Shallice’s SAS, the central executive functioned within cognitive psychology as a placeholder for unexplained control phenomena, lacking direct neuroanatomical correlates.

The transformation of working memory from a descriptive cognitive model into an empirically grounded cellular phenomenon was achieved through the pioneering neurophysiological research of Patricia Goldman-Rakic and her colleagues in the late 1970s and 1980s. Utilizing non-human primate models engaged in spatial delayed-response tasks, Goldman-Rakic inserted microelectrodes into the cortex surrounding the principal sulcus (Brodmann area 46) within the dorsolateral prefrontal cortex (dlPFC). She made the seminal discovery of delay-period activity: single prefrontal pyramidal neurons fired continuously during the brief delay interval separating the transient presentation of a spatial cue from the subsequent motor response, even though the physical cue was no longer visible in the environment.

This persistent neural firing across temporal gaps provided the first unequivocal neurobiological evidence of an internal mental representation bridging the gulf between perception and action. Goldman-Rakic demonstrated that this persistent activity was spatially tuned, resistant to distraction, and strictly predictive of behavioral accuracy; if the neuron’s firing faltered during the delay, the monkey inevitably made an error. However, Goldman-Rakic’s conceptualization initially framed prefrontal function predominantly as a short-term representational buffer—a high-fidelity storage container holding spatial coordinates in mind. Miller and Cohen recognized that while persistent delay activity demonstrated the prefrontal capacity for active maintenance, it did not fully explain executive control itself. The critical evolutionary leap of the prefrontal cortex was not merely its ability to store information over time, but its capacity to utilize those sustained internal representations as active, top-down control policies that systematically modify computation across sensory, limbic, and motor networks.

2. Core Architecture of the Guided Activation Theory (Miller & Cohen, 2001)

2.1 The Fundamental Thesis of Top-Down Biasing

The core thesis of the Guided Activation Theory, synthesized by Earl K. Miller and Jonathan D. Cohen in 2001, posits that cognitive control is fundamentally an act of competitive bias resolution. The brain is structurally organized as an interconnected web of processing pathways linking sensory inputs to behavioral outputs. Throughout an organism’s lifetime, repeated experiences, associative conditioning, and phylogenetic hardwiring sculpt synaptic weights within posterior and subcortical structures, rendering certain pathways exceedingly strong, rapid, and automatic. When sensory stimuli enter the nervous system, they naturally traverse these prepotent pathways to trigger default behavioral responses.

However, adaptive behavior frequently requires executing weak, unpracticed, or counter-intuitive responses in the presence of dominant, conflicting alternatives. The prefrontal cortex solves this problem not by generating motor acts de novo, nor by exerting direct brute-force suppression over subcortical reflexes, but by providing continuous, top-down modulatory biasing signals. Neurons within the PFC maintain active representations of behavioral goals, contextual contingencies, and rule structures. These prefrontal representations do not sit passively in a mnemonic silo; their axonal projections target intermediate and downstream processing units in sensory, association, and premotor structures.

By providing sustained, top-down excitatory drive to specific nodes along weaker processing streams, prefrontal biasing signals selectively enhance the responsiveness of those pathways. When an incoming sensory stimulus arrives, the pathway that has received prefrontal facilitation can successfully outcompete the default, prepotent pathway within local competitive networks. Miller and Cohen explicitly discarded the homuncular notion of an executive agent that makes executive decisions through unilateral commands. Instead, executive function emerges organically from the reciprocal, competitive dynamics established between feedforward sensory inputs and top-down modulatory representations. The prefrontal cortex serves as a contextual filter that tunes the internal conductance of the nervous system, allowing goal-relevant pathways to dominate motor execution.

2.2 Anatomical Connectivity Supporting Guided Activation

The structural feasibility of the Guided Activation Theory relies directly on the unique anatomical connectivity of the prefrontal cortex within the mammalian neuroaxis. Far from being an isolated computational module, the prefrontal cortex occupies an apex position within the cortical hierarchy, distinguished by its vast, reciprocal, polysynaptic connectivity with virtually every non-primary sensory, motor, and subcortical system.

Through dense reciprocal tracts, the lateral prefrontal cortex receives highly integrated feedforward signals from unimodal and multimodal association cortices. The dorsal visual stream (parietal cortex) projects spatial and kinematic information to the caudal and dorsal prefrontal regions, while the ventral visual stream (inferior temporal cortex) projects high-level object identities, faces, and semantic abstractions to the ventral prefrontal regions. Similarly, auditory association areas in the superior temporal gyrus and somatosensory association zones within the insula and parietal operculum converge within the lateral prefrontal mantle. This convergence creates a multidimensional representational zone where information across sensory modalities is unified into coherent, abstract representations of environmental states.

Simultaneously, the prefrontal cortex establishes bidirectional functional loops traversing the fronto-striatal-thalamic circuitry. Descending corticostriatal projections terminate densely in the caudate nucleus and anterior putamen, modulating basal ganglia gating circuits that regulate the release of motor programs and the updating of internal working memory buffers. In return, the prefrontal cortex sends massive, divergent top-down projections cascading back into the intermediate and early stages of sensory processing hierarchies, including the secondary visual areas, sensory association cortices, and premotor structures. These descending feedback projections terminate primarily on pyramidal neurons and local inhibitory interneurons, providing the exact physical anatomical substrate required to deliver modulatory biases that selectively amplify signal-to-noise ratios in posterior processing streams.

2.3 The Metaphor of the Railway Switchyard

To provide an intuitive conceptual model of their neurocomputational framework, Miller and Cohen introduced the evocative metaphor of the railway switchyard. In this mechanical analogy, the neural pathways traversing sensory inputs to motor effectors are conceptualized as physical railroad tracks laid across the landscape of the brain. High-speed, heavy-gauge tracks represent strongly consolidated, habitual, and prepotent stimulus-response routines—such as looking toward a sudden flash of light, reading printed text upon visual exposure, or grasping an object presented to the hand. Signals entering the brain travel down these default tracks automatically, propelled by the high kinetic momentum of overlearned synaptic connectivity.

Under routine conditions, trains (sensory-driven neural signals) glide smoothly along these primary tracks without requiring active intervention. However, dynamic and rule-governed environments frequently present situations where a train must be routed away from the main track onto an infrequently used, low-speed siding—representing a weak, novel, or effortful action, such as looking away from a sudden flash (antisaccade) or naming the ink color of an incongruent color-word rather than reading the word itself. If the system relies entirely on the tracks’ existing geometry, the train will inevitably follow the main line, resulting in an error of habit.

Within this metaphor, the prefrontal cortex does not construct a completely new set of tracks, nor does it physically carry the train to its destination. Instead, prefrontal representations act as the mechanical switches situated at track junctions. By setting and maintaining the position of the switch prior to the train’s arrival, prefrontal biasing signals alter the physical alignment of the junction. When the sensory signal arrives, the prefrontal switch redirects the neural activation pattern onto the weaker, goal-appropriate branch line. If prefrontal function is damaged, absent, or transiently disrupted, the switches revert to their default positions; the train invariably hurtles down the main line, yielding classical clinical symptoms of environmental dependency, utilization behavior, and perseveration.

3. Top-Down Biasing and the Mechanics of Neural Competition

3.1 Biased Competition Principles in Sensory and Motor Cortices

The neurocomputational engine powering Guided Activation is rooted directly in the Biased Competition Model originally formulated by Robert Desimone and John Duncan (1995) to explain visual selective attention. Desimone and Duncan recognized that the nervous system is fundamentally capacity-limited: the massive sensory influx entering through the sensory receptors far exceeds the processing and motor-execution bandwidth of downstream structures. Consequently, stimuli present within the visual field must engage in continuous, parallel competition for representation, cognitive resources, and behavioral access.

In sensory cortices—such as intermediate visual area V4 and higher-level inferior temporal (IT) cortex—individual neurons possess receptive fields containing multiple visual objects. When two distinct stimuli (e.g., a green apple and a red ball) fall simultaneously within the same neuronal receptive field, they do not evoke an additive sum of their independent responses. Instead, they mutually suppress one another through local, lateral inhibitory interactions mediated by cortical interneuron networks. If left unguided, the stimulus with greater physical salience (e.g., higher luminance contrast, rapid motion, or sudden onset) dominates the competition, commanding the neuron’s firing rate and effectively filtering out the less salient object.

Top-down biasing provides the counterweight to bottom-up sensory dominance. When an organism searches for a specific target (e.g., searching for a red object), the prefrontal cortex and associated frontoparietal structures generate an excitatory bias directed toward neurons that process the target’s defining features. This modest top-down excitatory drive acts as a lever within the local competitive circuit. By artificially elevating the baseline excitability of target-congruent neurons, the prefrontal signal tips the balance of mutual inhibition. The favored neuronal pool amplifies its firing, which in turn delivers robust lateral inhibition to suppress neighboring neuronal assemblies encoding distractor features. Miller and Cohen realized that this biased competition mechanism was not confined to sensory visual processing; it operates universally across associative, linguistic, emotional, and motor selection networks.

3.2 Interaction of Bottom-Up Salience and Top-Down Intentionality

The behavioral state of an organism at any given millisecond reflects a dynamic equilibrium between two opposing forces: bottom-up salience (exogenous, stimulus-driven capture) and top-down intentionality (endogenous, goal-directed bias). Bottom-up processing is intrinsically rapid, parallel, and driven purely by the physical attributes of incoming stimuli relative to their immediate background. A sudden loud noise, a bright flash of light, or a rapidly moving predator triggers feedforward bursts of action potentials that race unimpeded through sensory hierarchies into subcortical reflexive hubs, such as the superior colliculus and the amygdala.

In contrast, top-down biasing signals are comparatively slow, deliberate, and metabolically expensive. The prefrontal cortex must retrieve, configure, and actively sustain the appropriate rule set or goal state before its descending axonal projections can alter downstream synaptic conductances. In formal mathematical representations of connectionist networks, the net input ($I_i$) driving a given processing unit ($i$) within an intermediate or motor network can be formalized as an additive function of bottom-up feedforward inputs, lateral competitive interactions, and prefrontal modulatory bias:

$$I_i = \sum_{j} W_{ji} X_j + \sum_{k} L_{ki} Y_k + \beta B_i$$

In this simplified formulation, $W_{ji}$ represents the synaptic weights of the feedforward sensory inputs $X_j$, $L_{ki}$ denotes the inhibitory lateral connections from competing local units $Y_k$, and $B_i$ represents the top-down biasing signal emanating from prefrontal goal units, scaled by a modulatory gain factor $\beta$.

Because feedforward weights ($W_{ji}$) are often highly optimized through evolution and extensive repetitive learning, the sensory burst arrives first, commanding early processing. For top-down intentionality to triumph, the prefrontal biasing signal ($\beta B_i$) must arrive in time and possess sufficient amplitude to offset the sensory-driven momentum before the motor threshold is breached. If the prefrontal bias arrives late, or if its amplitude fluctuates due to cognitive fatigue, distraction, or neurochemical depletion, bottom-up salience commandeers the motor apparatus, resulting in an attentional capture or an erroneous habitual response.

3.3 Inhibition via Selective Excitatory Amplification

One of the most persistent misconceptions in classical cognitive neurology is the assumption that the prefrontal cortex exerts behavioral control through direct, long-range inhibitory projections that clamp down on lower brain regions. Neuroanatomical investigations have definitively demonstrated that the vast majority of long-range projection neurons in the mammalian neocortex are glutamatergic, which means they are exclusively excitatory. Prefrontal projection neurons cannot directly hyperpolarize distant target neurons in the temporal lobes, parietal lobes, or motor execution circuits.

The Guided Activation Theory elegantly resolved this physiological paradox: prefrontal inhibition is an emergent, network-level phenomenon achieved via selective excitatory amplification. When the prefrontal cortex acts to suppress a prepotent habit, it does not send direct inhibitory signals to the motor programs encoding that habit. Instead, prefrontal glutamatergic efferents selectively target and excite the specific subset of neurons representing the weak, goal-appropriate alternative.

Once these goal-relevant neurons receive prefrontal excitation, they engage the local microcircuitry of the target area. All cortical and subcortical processing areas are saturated with networks of local, highly arborized GABAergic interneurons (such as parvalbumin-positive basket cells). As the prefrontal-favored neurons increase their firing rates, they simultaneously recruit these local inhibitory interneurons. The interneurons spread a wide blanket of lateral inhibition across the immediate cortical neighborhood, effectively suppressing the alternative, competing assemblies—including the prepotent, automatic pathways. Consequently, apparent prefrontal “inhibition” is completely mediated by top-down excitation paired with local competitive dynamics. The prefrontal cortex selects what to activate; the intrinsic biophysics of downstream competitive networks handles the necessary suppression.

4. Representation of Goals, Rules, and Behavioral Context in the PFC

4.1 Encoding Abstract Rules and Action Policies

For top-down biasing signals to guide processing effectively across diverse behavioral scenarios, the prefrontal cortex must represent information at a high level of abstraction. If prefrontal neurons were merely sensitive to raw sensory parameters (such as wavelengths of light or acoustic frequencies), they would be incapable of coordinating behavior across varying contexts where completely different physical stimuli demand identical functional responses. Pioneering non-human primate studies conducted by Earl K. Miller and colleagues in the late 1990s and early 2000s provided the first direct electrophysiological proof that single prefrontal neurons encode abstract, categorical rules independent of physical stimulus attributes.

In classic experiments, monkeys were trained to execute matching tasks governed by abstract rules: in some trials, they were required to perform a “same” (delayed match-to-sample) rule, while in other trials they had to perform a “different” (delayed non-match-to-sample) rule, signaled by arbitrary visual cues. Crucially, the sensory stimuli used to test the rule (pictures of animals, abstract shapes, complex textures) varied continuously from trial to trial. Microelectrode recordings revealed that a significant subpopulation of lateral prefrontal neurons fired vigorously whenever the abstract rule was “same,” regardless of which specific cue signaled the rule, and irrespective of which physical images were being matched. Other neurons fired selectively for the “different” rule.

This abstraction of representation extends hierarchically across multiple cognitive domains:

• First-order stimulus-response bindings: Simple associations directly linking a sensory feature to a specific motor act (e.g., Red $\rightarrow$ Press Left, Green $\rightarrow$ Press Right).
• Contextual contingencies: Conditional rules dictating how first-order bindings must be interpreted based on an environmental cue (e.g., If Tone A sounds, Red $\rightarrow$ Left; if Tone B sounds, Red $\rightarrow$ Right).
• Abstract relational schemas: Generalized policies operating over relational concepts, such as “greater than,” “match vs. non-match,” or “maximize social reciprocity.”

Prefrontal neurons stabilize these abstract contextual policies by forming recurrent, self-sustaining attractor states within local microcircuits, establishing an internal frame of reference that protects current task sets against sensory distraction.

4.2 Representational Flexibility and Adaptive Coding

The prefrontal cortex does not dedicate rigid, immutable populations of neurons to single, pre-determined tasks. Such a static organizational scheme would rapidly exhaust the brain’s finite cellular real estate, leaving the organism incapable of learning novel tasks or navigating unfamiliar environments. In parallel with the formulation of Guided Activation, John Duncan proposed the Adaptive Coding Model of prefrontal function, which directly complements Miller and Cohen’s architectural principles.

Duncan demonstrated that prefrontal cortical neurons exhibit extraordinary functional plasticity, dynamically reallocating their computational resources to reflect the specific demands, contingencies, and dimensionalities of the ongoing task. An individual prefrontal neuron that exhibits sharp selective tuning for spatial coordinates in a spatial navigation task can smoothly reconfigure its firing properties within minutes to encode color categories, semantic identities, or abstract reward values when the animal transitions to a visual categorization paradigm.

This immense representational flexibility is driven computationally by nonlinear mixed selectivity, a phenomenon rigorously documented by Mattia Rigotti, Stefano Fusi, and colleagues (2013). Rather than encoding behavioral variables in an isolated, linear, one-by-one fashion, individual prefrontal neurons simultaneously encode complex, nonlinear combinations of multiple task variables—such as stimulus identity, rule context, motor preparation, and reward expectation. At the population level, mixed selectivity embeds representations within a massive, high-dimensional neural space. This high dimensionality is computationally advantageous: it allows simple linear readouts by downstream motor systems to instantly categorize complex behavioral states and discriminate between subtly different rule contingencies, conferring near-infinite flexibility upon prefrontal control policies.

4.3 Maintenance versus Utilization of Contextual Information

A foundational theoretical insight of the Guided Activation Theory is the sharp functional dissociation between the passive maintenance of information in working memory and the active utilization of contextual representations to drive cognitive control. Holding an item in a short-term representational buffer is a necessary, but entirely insufficient, condition for executive function. For instance, an individual can easily remember a phone number while walking into a room, yet completely forget why they entered that room; the numerical digits were maintained in working memory, but the contextual policy guiding behavioral action was lost or uncoupled from motor selection.

Active maintenance relies on biophysical mechanisms such as recurrent excitatory collateral connections among prefrontal pyramidal neurons, slow N-methyl-D-aspartate (NMDA) receptor kinetics, and short-term synaptic facilitation mediated by intracellular calcium microdomains. These mechanisms maintain an elevated baseline firing rate or preserve a synaptic memory trace across temporal intervals.

However, utilization requires that these persistently firing prefrontal populations project their activity down into the rest of the brain to operate as active filters. The maintained representation must function as a top-down template:

• It projects to early visual cortices, priming feature detectors matching the target goal.
• It projects to basal ganglia loops, gating the execution of incompatible motor programs.
• It alters synaptic gain within posterior association hubs, suppressing distracting sensory streams.

This utilization process is extraordinarily sensitive to external disruptions. Under high cognitive loads, during the attentional blink, or under acute psychosocial stress, prefrontal context representations may remain physically intact within their local attractors, but their top-down biasing efficacy degrades catastrophically. When this decoupling occurs, the organism defaults to reflexive, stimulus-driven behaviors, exhibiting goal neglect despite preserving the conscious, declarative knowledge of what they were supposed to do.

5. Computational Modeling: Connectionist Frameworks and Neural Implementations

5.1 The Cohen, Dunbar, and McClelland (1990) Stroop Model

The computational engine that gave birth to the Guided Activation Theory was the classic parallel distributed processing (PDP) model of the Stroop task published by Jonathan D. Cohen, Kevin Dunbar, and James L. McClelland in 1990. The Stroop effect represents the archetypal laboratory paradigm for testing cognitive control: when participants view color words (e.g., the word “RED”) printed in conflicting colored inks (e.g., green ink) and are instructed to name the ink color, they exhibit severe reaction time slowing and elevated error rates. Conversely, when instructed to read the word, ink color causes virtually zero interference.

Cohen, Dunbar, and McClelland modeled this empirical phenomenon using a feedforward neural network equipped with competing pathways. The network possessed input layers for color features and word features, an intermediate layer representing hidden processing units, and an output layer representing speech production units. The resting synaptic connection weights between the word input units and the output units were set to high values, simulating years of reading experience and heavy overlearning. In contrast, the synaptic connection weights along the color-naming pathway were set to low values, reflecting the weaker, less-practiced nature of naming ink colors.

To implement task control, the authors introduced Task Demand Units, which served as the conceptual prototype for the prefrontal cortex in the Guided Activation Theory. When the model was instructed to name the ink color, the corresponding Task Demand Unit was activated, injecting continuous, additive top-down bias into the intermediate units of the color-naming pathway. This additive bias elevated the intermediate color units’ activations along their sigmoidal activation functions, pushing them into a higher gain state. As a result, activation swept down the weaker pathway fast enough to outcompete the word pathway at the output layer. The model flawlessly replicated the classic asymmetric interference of the Stroop effect, confirmed that interference occurs continuously rather than in discrete, all-or-nothing stages, and demonstrated mathematically that cognitive control does not require an intelligent homunculus, but simply an additive source of sustained bias.

5.2 Biophysically Realistic Attractor Network Models

Following the abstract connectionist success of the Cohen et al. (1990) model, computational neuroscientists sought to implement guided activation within biophysically realistic, spiking neural networks. Spearheaded by Xiao-Jing Wang and colleagues, these models abandoned idealized connectionist nodes in favor of integrate-and-fire or conductance-based pyramidal neurons and GABAergic interneurons governed by realistic membrane biophysics.

Central to these biophysical implementations is the critical reliance on NMDA receptor channels versus AMPA receptor channels within the prefrontal microcircuitry. Fast sensory inputs and motor commands in posterior cortices are mediated predominantly by low-latency AMPA and $GABA_A$ receptors, which turn on and off within milliseconds. However, persistent attractor states within the prefrontal cortex require sustained firing that resists rapid decay and internal fluctuations. NMDA receptors exhibit uniquely slow unbinding kinetics (with decay time constants on the order of 50 to 100 milliseconds) and a voltage-dependent magnesium ($Mg^{2+}$) block that acts as a coincidence detector.

Computational simulations demonstrate that prefrontal networks dominated by recurrent NMDA-mediated excitatory connections among pyramidal cells can generate stable, self-sustaining attractor states. Once triggered by an environmental cue or an internal decision, a specific population of prefrontal neurons remains in a persistent high-firing state for seconds or minutes. Simultaneously, recurrent projections to local parvalbumin-positive $GABA_A$ interneurons form a competitive inhibitory basin around the attractor. This prevents rival populations from firing, effectively insulating the active rule representation against distraction. When the task demands a rule switch, a transient inhibitory pulse or a neuromodulatory surge destabilizes the active attractor, permitting the network to transition smoothly through phase-space into an alternative, goal-appropriate attractor basin.

5.3 Reinforcement Learning and Synaptic Plasticity in Control Models

A critical question left partially unanswered by the original 2001 Guided Activation formulation was: how do prefrontal task demand units learn which rules to activate in the first place? To address this, subsequent computational architectures integrated guided activation with formal reinforcement learning (RL) theory and actor-critic computational architectures.

In these integrated models, the prefrontal cortex functions as an adaptive “actor” maintaining context, while the basal ganglia and dopaminergic midbrain act as a “critic” computing temporal difference reward prediction errors (RPEs):

$$\delta(t) = r(t) + \gamma V(S_{t+1}) – V(S_t)$$

Here, $\delta(t)$ represents the dopamine prediction error, $r(t)$ is the immediate reward, $\gamma$ is the discount factor, and $V(S)$ represents the expected state value. When an organism executes a novel rule configuration and receives an unexpected reward, phasic bursts of dopamine are released across the prefrontal cortex and the striatum. In the lateral prefrontal cortex, this dopamine surge triggers synaptic plasticity governed by modified Hebbian rules:

$$\Delta W_{ij} = \eta \cdot \delta(t) \cdot Pre_i \cdot Post_j$$

Through successive reward prediction errors, the synaptic weights connecting sensory context cues to specific prefrontal rule representations are systematically strengthened. Furthermore, reinforcement learning models demonstrate how gating mechanisms learn to shield prefrontal representations during delay intervals and selectively ungate them precisely when feedback indicates that task contingencies have shifted, achieving fully autonomous, unsupervised optimization of cognitive control networks.

6. Empirical Neurophysiological Evidence in Non-Human Primates

6.1 Single-Unit and Multi-Unit Recordings in Lateral PFC

The neurobiological reality of the Guided Activation Theory has received profound empirical confirmation from single-unit and multi-unit electrophysiology in awake, behaving rhesus macaques (Macaca mulatta). The laboratory of Earl K. Miller has systematically designed paradigms explicitly aimed at dissociating pure sensory processing, abstract rule selection, and motor execution.

In a series of landmark investigations, David J. Freedman, Earl K. Miller, and colleagues trained monkeys to categorize computer-generated visual stimuli morphed systematically across a continuum between “cats” and “dogs.” The animals were required to indicate whether a stimulus belonged to the cat category or the dog category, regardless of its specific blend percentage (e.g., a morph comprising 60% cat and 40% dog was rewarded as a cat). Microelectrode recordings in the lateral prefrontal cortex demonstrated that prefrontal neurons did not track the continuous physical changes in the images. Instead, their firing rates exhibited sharp, categorical tuning curves: they responded robustly and uniformly to all stimuli along the “cat” side of the boundary, and showed an abrupt, step-like suppression of activity the moment a stimulus crossed the 50% category threshold to the “dog” side.

Simultaneous recording across dozens of prefrontal electrodes using high-density arrays has allowed the reconstruction of dynamic population vectors. These multi-unit analyses show that prefrontal neural populations do not fire in isolation; they coordinate their activity into collective trajectories within a lower-dimensional state space. When rule contingencies change—such as when a monkey is signaled to switch from a “match by color” rule to a “match by shape” rule—the prefrontal population trajectory undergoes an immediate, high-velocity reorganization. This population-level state transition occurs hundreds of milliseconds prior to the corresponding re-tuning of sensory processing in extrastriate cortex, providing undeniable electrophysiological evidence that prefrontal populations configure the downstream processing landscape in advance of action.

6.2 Local Field Potentials, Coherence, and Inter-Area Synchrony

While action potentials represent the primary units of computation within local circuits, long-range communication between the prefrontal cortex and distant posterior structures is organized through local field potentials (LFPs) and oscillatory synchrony. The Guided Activation Theory has found an indispensable neurophysiological partner in the Communication-Through-Coherence (CTC) hypothesis formulated by Pascal Fries.

Empirical recordings reveal that when top-down prefrontal biasing is actively engaged, distinct frequency bands coordinate feedforward versus feedback information flow across the cortical hierarchy:

• Gamma rhythms (>30–80 Hz): Associated with localized, bottom-up sensory processing and feedforward information flow traversing superficial cortical layers (layers II/III).
• Beta rhythms (13–30 Hz): Predominantly generated within deep cortical layers (layers V/VI) and specialized for top-down cognitive control, rule maintenance, and the stabilization of the behavioral status quo.
• Theta/Alpha rhythms (4–12 Hz): Serve as long-range phase-modulators that periodically gate the excitability of sensory cortices, organizing processing into discrete temporal windows.

Groundbreaking studies by Buschman and Miller (2007, 2012) demonstrated that during top-down, goal-directed search tasks, coherent beta-band oscillations synchronize the lateral prefrontal cortex with the posterior parietal cortex. Prefrontal beta oscillations lead the parietal cortex in time, effectively entraining parietal circuits into phase alignment. This rhythmic alignment ensures that when prefrontal biasing spikes arrive at downstream dendritic trees, they hit the receptive neurons precisely during their open, depolarized excitability peaks. Conversely, when visual search is driven by bottom-up salience (a bright pop-out target), the parietal cortex exhibits strong gamma-band synchrony that leads the prefrontal cortex in time. Guided activation is therefore dynamic, flexible, and structurally paced by coherent inter-areal rhythms.

6.3 Causal Manipulations: Lesioning, Cooling, and Optogenetics

Correlational electrophysiology cannot establish whether prefrontal delay-period firing and beta synchrony are strictly causal for behavioral control. To confirm causality, neuroscientists have utilized targeted reversible and permanent manipulations, including cryogenic cooling, focal pharmacological inactivation, electrical microstimulation, and optogenetics.

Reversible cryogenic cooling of the dorsolateral prefrontal cortex in primates provides an exceptional experimental tool: by circulating chilled fluid through implanted cryoloops, the temperature of localized cortical tissue can be brought below 20°C, selectively silencing synaptic transmission without interrupting passing axonal fibers. When dlPFC cryoloops are activated, monkeys experience immediate, severe deficits in tasks requiring top-down guided activation (such as the delayed non-match-to-sample task and spatial delayed response). Crucially, the animals can still execute basic sensory discrimination tasks and reach accurately toward illuminated visual targets, proving that cryogenic silencing does not disrupt sensory reception or motor execution per se; it selectively abolishes the top-down biasing signals required to resolve ambiguous contingencies.

In a classic study demonstrating the specific biasing mechanics of frontal control, Tirin Moore and Katherine Armstrong (2003) applied subthreshold electrical microstimulation to the frontal eye fields (FEF)—an anterior oculomotor region structurally linked to prefrontal control circuits. The authors lowered microelectrodes into visual area V4 to record the sensory receptive fields of single neurons while simultaneously delivering electrical microcurrents below the threshold for evoking a saccadic eye movement into the corresponding FEF site. When subthreshold microstimulation was applied to the FEF, the sensory response of neurons in area V4 increased significantly; the microstimulation amplified the neural response to a visual target placed in the neuron’s receptive field, mimicking the exact neurophysiological signature of natural, voluntary selective attention. Modern optogenetic studies in transgenic rodents have extended these findings by selectively expressing channelrhodopsin (ChR2) in prefrontal cortico-striatal and cortico-sensory projection neurons. Illuminating these specific prefrontal terminals directly enhances target discrimination and suppresses impulsive default reactions, confirming the biophysical mechanics of guided activation at the single-pathway level.

7. Human Neuroimaging and Neuropsychological Validations

7.1 fMRI Investigations of the Stroop Task and Attentional Conflict

The advent of functional Magnetic Resonance Imaging (fMRI) in the 1990s enabled neuroscientists to test the predictions of the Guided Activation Theory directly within the human brain. The classic Stroop task, along with its conceptual derivatives (such as the Flanker task and the Simon task), provided the primary empirical testing ground.

Consistent with the predictions of Miller and Cohen, fMRI studies reliably demonstrate that when human participants face high-conflict trials—such as color-word incongruency—there is a dramatic increase in the blood-oxygen-level-dependent (BOLD) signal across a distributed frontoparietal cognitive control network, with prominent nodes in the bilateral dorsolateral prefrontal cortex (dlPFC, middle frontal gyrus) and the dorsal anterior cingulate cortex (dACC). Advanced connectivity analyses, such as Dynamic Causal Modeling (DCM) and psychophysiological interactions (PPI), have unequivocally established that effective connectivity flows directionally from the dlPFC to posterior sensory processing zones. For example, during tasks requiring selective attention to faces versus houses in visual arrays, dlPFC activation causally modulates the BOLD gain in the fusiform face area (FFA) and the parahippocampal place area (PPA), respectively.

Furthermore, event-related fMRI designs have successfully isolated preparatory, top-down biasing activity from subsequent conflict resolution. When a visual cue informs human participants of an upcoming task switch (e.g., preparing to name colors instead of reading words), the dlPFC exhibits robust preparatory BOLD elevation *prior* to the onset of the actual stimulus. The magnitude of this preparatory dlPFC activation predicts both the degree of conflict reduction on the subsequent trial and the speed of the participant’s reaction time. Multi-voxel pattern analysis (MVPA) has taken these findings a step further: machine-learning classifiers trained on multi-voxel BOLD patterns within the lateral prefrontal cortex can decode which specific rule, task set, or behavioral context an individual is holding in mind, long before the motor response is generated.

7.2 Human Lesion Studies and Dysexecutive Syndrome

Human clinical neuropsychology provides compelling naturalistic validations of the Guided Activation Theory through the study of patients suffering from focal prefrontal cortex lesions resulting from cerebrovascular accidents, traumatic brain injury, or surgical resections.

The classic neuropsychological benchmark for prefrontal damage is the Wisconsin Card Sorting Test (WCST). In this task, patients must sort cards according to changing sorting rules: color, shape, or number. The sorting rule is never explicitly stated; the patient must infer it from categorical feedback (“correct” or “incorrect”). Once the patient identifies the rule and sorts ten consecutive cards accurately, the sorting rule changes without warning. Neurologically intact individuals rapidly notice the change after an error, suppress the old rule, and systematically test new rules. In stark contrast, patients with bilateral or dorsolateral prefrontal lesions exhibit pronounced perseveration: they continue sorting cards according to the old, previously reinforced rule, trial after trial, despite receiving continuous verbal feedback that their choices are completely wrong.

Under the Guided Activation Theory, perseveration is not an impairment of perceptual discrimination or declarative comprehension; it is a direct mechanical consequence of failing to generate an endogenous biasing signal. The old rule, reinforced over multiple preceding trials, has established temporary dominance within the posterior associative network. When the contingency shifts, the posterior pathway remains prepotent. Without an intact prefrontal cortex to inject an excitatory bias into the alternative, unreinforced rule pathways, the patient cannot overcome the momentum of the previously established habit. This explains the classic clinical phenomenon of the “knowing-doing dissociation”: frontal patients will frequently look at the experimenter and explicitly declare, “This is wrong, but I have to do it,” sorting the card incorrectly even as they speak. Without prefrontal guided activation, conscious awareness cannot bridge into motor execution.

At the most extreme end of prefrontal pathology lies utilization behavior and environmental dependency syndrome, first characterized in detail by François Lhermitte. Patients with extensive bifrontal pathology become absolute slaves to bottom-up perceptual affordances. If an examiner places a pair of eyeglasses on a table in front of such a patient, the patient will immediately pick them up and put them on, even if they are already wearing their own glasses. If presented with a glass of water and a toothbrush, they will immediately brush their teeth. In these patients, the high-speed railway tracks of sensory-to-motor processing are entirely intact, but the prefrontal switches have been completely destroyed. Every environmental stimulus that hits their sensory receptors triggers its default, overlearned motor routine without any regulatory mediation from internal goals.

7.3 Electrophysiological Markers (ERP and MEG) in Humans

The temporal dynamics of guided activation operate on the timescale of tens to hundreds of milliseconds, demanding the millisecond temporal resolution of scalp-recorded event-related potentials (ERPs), magnetoencephalography (MEG), and direct intracranial electroencephalography (iEEG).

Decades of electrophysiological research have identified specific ERP waveforms that track the detection of conflict and the deployment of prefrontal control:

• The N200 (or Frontocentral N2): A negative-going deflection peaking between 200 and 350 milliseconds post-stimulus over medial-frontal electrodes. The amplitude of the N200 scales directly with the magnitude of conflict between competing response pathways (e.g., Flanker incongruent vs. congruent trials), indexing the transient call for control.
• The Error-Related Negativity (ERN): A sharp negative deflection occurring within 0 to 100 milliseconds *after* an erroneous motor execution, generated predominantly within the dorsal anterior cingulate cortex, signaling a mismatch between the intended prefrontal goal state and actual motor output.
• The P300 Complex (P3a and P3b): The P3a represents a frontocentral positive deflection associated with the involuntary capture of attention by novel or unexpected stimuli, while the P3b reflects a slower, centroparietal positivity indexing the top-down updating of context in working memory and the cognitive closure of a task set.

High-density MEG investigations have successfully tracked the millisecond-scale propagation of top-down biasing signals across human cortex. When an unexpected cue triggers an urgent task switch, MEG reveals a rapid, phase-locked burst of low-frequency oscillations in the dorsolateral prefrontal cortex within 120 milliseconds. This frontal burst is followed roughly 80 milliseconds later by an amplification of evoked magnetic fields in ventral temporal and visual areas, accompanied by a targeted suppression of alpha-band (8–12 Hz) power over task-relevant sensory representations. Intracranial recordings from human patients undergoing surgical monitoring for drug-resistant epilepsy have confirmed these findings: high-gamma power (70–150 Hz)—a direct proxy for local multi-unit spiking—elevates across the lateral prefrontal cortex first, establishing phase-locking with downstream sensory structures before the behavioral response is initiated.

8. Interactions with the Basal Ganglia and Neuromodulatory Systems (Dopamine)

8.1 The Basal Ganglia Gating Hypothesis (PBWM Model)

While the original Guided Activation Theory primarily emphasized prefrontal top-down outputs to posterior cortical regions, the prefrontal cortex does not operate in an anatomical vacuum. Its functional integrity depends fundamentally on closed, parallel, loop-like circuits connecting it with the subcortical basal ganglia. To explain how prefrontal representations are dynamically updated and shielded from unwanted interference, Michael J. Frank, Randall C. O’Reilly, and colleagues developed the Prefrontal Cortex Basal Ganglia Working Memory (PBWM) computational architecture.

The PBWM model frames the basal ganglia as a dynamic, intelligent gating mechanism for prefrontal working memory. A major computational dilemma confronting any cognitive control system is the stability-flexibility trade-off:

• The Stability Requirement: Once a goal or rule representation is loaded into the prefrontal cortex, it must be vigorously defended and shielded against distracting, irrelevant perceptual noise.
• The Flexibility Requirement: The moment environmental contingencies shift or a goal is satisfied, the prefrontal network must rapidly discard the obsolete representation and update its attractors with new information.

The PBWM model solves this trade-off through a division of labor between the prefrontal cortex and the striatal direct and indirect pathways. Under baseline conditions, the output structures of the basal ganglia (the internal segment of the globus pallidus, GPi, and the substantia nigra pars reticulata, SNr) exert tonic, high-frequency GABAergic inhibition onto the thalamic nuclei projecting to the prefrontal cortex. This tonic inhibition keeps the “gate” closed, effectively locking the prefrontal cortex into a stable, recurrent attractor state that ignores environmental distractions.

When an environmental cue indicates that an update is mandatory, the striatal direct pathway is activated. Striatal medium spiny neurons (MSNs) fire, transiently inhibiting the GPi/SNr. This action produces disinhibition of the thalamus: the thalamic gate bursts open, allowing a strong wave of feedforward sensory information to flood into the prefrontal cortex. Once the new goal state is successfully transferred into prefrontal recurrent attractors, the striatal indirect pathway reasserts dominance, closing the thalamic gate once more and locking the new representation in place. Thus, the basal ganglia control the gating of working memory, while the prefrontal cortex maintains and executes the top-down guided activation.

8.2 Dopaminergic Modulation of Prefrontal Signal-to-Noise Ratio

The biophysical mechanics of guided activation within prefrontal microcircuits are profoundly modulated by ascending monoaminergic neurotransmitters, chief among which is dopamine. Mesocortical dopaminergic projections originate in the ventral tegmental area (VTA, A10 cell group) and terminate heavily throughout the prefrontal layers, where they act upon two primary classes of G-protein-coupled receptors: the $D_1$-like receptor family ($D_1$ and $D_5$, coupled to $G_{\alpha s}$) and the $D_2$-like receptor family ($D_2$, $D_3$, and $D_4$, coupled to $G_{\alpha i/o}$).

Decades of pharmacological and electrophysiological investigations, particularly those conducted by Patricia Goldman-Rakic and Amy Arnsten, have demonstrated that the relationship between prefrontal $D_1$ receptor stimulation and cognitive control performance follows an inverted-U shaped dose-response curve. Insufficient dopamine release (as observed in unmedicated ADHD or chronic stress exhaustion) or excessive dopamine flooding (as observed during intense acute traumatic stress) causes severe degradations in executive performance:

$$Performance = f\left(\frac{1}{1 + e^{-k(D_1 – \theta_1)}} – \frac{1}{1 + e^{-k(D_1 – \theta_2)}}\right)$$

At the cellular level, the $D_1$ and $D_2$ receptor families play sharply distinct, complementary roles in regulating guided activation:

• D1 Receptor Stimulation: Promotes attractor stability. $D_1$ activation increases intracellular cyclic adenosine monophosphate (cAMP) and activates protein kinase A (PKA), selectively amplifying slow NMDA receptor-mediated currents while dampening spontaneous, non-specific background AMPA currents. This process increases the signal-to-noise ratio (SNR) of the prefrontal network, deepening the attractor basin of the currently active goal and rendering it immune to distractor interference.
• D2 Receptor Stimulation: Promotes cognitive flexibility and state transitions. $D_2$ receptor activation suppresses adenylate cyclase, reducing NMDA conductance and lowering the energy barriers between competing attractor states. This action allows the network to transition fluidly away from an obsolete rule, enabling rapid task-switching and cognitive re-tuning.

8.3 Noradrenergic and Cholinergic Influences on Cognitive Control

Beyond dopamine, the execution of top-down guided activation requires precise neuromodulatory orchestration by the ascending noradrenergic and cholinergic systems.

The locus coeruleus-norepinephrine (LC-NE) system exerts widespread influence over prefrontal computational dynamics. In their seminal Adaptive Gain Theory, Gary Aston-Jones and Jonathan D. Cohen (2005) proposed that the LC-NE system functions in two distinct operational modes:

• Phasic Mode: Characterized by low baseline tonic firing coupled to sharp, high-amplitude phasic bursts of norepinephrine released in response to task-relevant decision events. Phasic norepinephrine acts upon high-affinity post-synaptic $\alpha_2A$-adrenoceptors in the prefrontal cortex, enhancing synaptic connectivity within task-related pyramidal ensembles. This state optimizes the “exploit” mode of behavior, locking the brain into focused, goal-directed guided activation.
• Tonic Mode: Characterized by elevated baseline firing without phasic bursting, occurring under conditions of prolonged unrewarding outcomes or high environmental uncertainty. High levels of norepinephrine flood lower-affinity $\alpha_1$ and $\beta$-adrenoceptors, disrupting prefrontal attractors and producing distractibility. This state optimizes the “explore” mode of behavior, prompting the organism to disengage from its current task and search for alternative environmental opportunities.

Simultaneously, the ascending basal forebrain cholinergic system delivers targeted pulses of acetylcholine (ACh) across both sensory and prefrontal cortices. Groundbreaking work by Michael Hasselmo demonstrated that high cholinergic tone acts as an internal functional switch: acetylcholine stimulates nicotinic and $M_1$ muscarinic receptors on feedforward sensory fibers, significantly boosting sensory-driven bottom-up transmission while suppressing internal, recurrent excitatory feedback mediated by $M_2/M_4$ auto-receptors. High acetylcholine tone enables the rapid acquisition of novel environmental data, whereas low acetylcholine tone allows internal prefrontal models and guided activation biases to dominate processing without continuous sensory disruption.

9. Cognitive Flexibility, Task Switching, and Interference Resolution

9.1 The Mechanisms of Task Switching and Switch Costs

The operational resilience of guided activation is subjected to its most stringent empirical test within task-switching paradigms. In these experiments, human participants alternate between two or more distinct task sets operating on identical multidimensional stimuli (for example, categorizing a number-letter pair such as “7G” either as odd/even or as consonant/vowel depending on a preceding instructional cue). When individuals switch from one task to another, they consistently exhibit a switch cost: a measurable elevation in reaction times and error rates relative to trials where the task is repeated.

Under the Guided Activation Theory, total switch costs deconstruct into two computationally distinct phenomena:

• Endogenous Task-Set Reconfiguration: The active, top-down process driven by the prefrontal cortex to clear the obsolete rule representation, configure the new task demand units, and propagate biasing signals down to the sensory and motor effectors. This component is time-dependent: if participants are provided with a long cue-to-stimulus interval (CSI), they can prepare in advance, significantly attenuating the switch cost.
• Exogenous Task-Set Inertia: The passive, bottom-up carryover effects persisting within intermediate and motor processing pathways. Synaptic trace facilitation, refractory interneuron dynamics, and lingering associative weights from the preceding trial continue to resist the new rule.

A profound validation of the Guided Activation Theory is its ability to explain the counter-intuitive phenomenon of asymmetric switch costs. When participants switch between an easy, dominant task (e.g., word reading) and a difficult, non-dominant task (e.g., color naming), the cost of switching to the easy task is paradoxically much higher than the cost of switching to the hard task. Miller and Cohen’s model explains this with mechanical precision: performing the difficult task requires the prefrontal cortex to inject an exceptionally massive top-down bias to overcome the dominant habit. When the task suddenly switches back to the easy habit on the subsequent trial, that colossal prefrontal bias cannot be instantaneous extinguished; it lingers as intense proactive interference, actively fighting against the execution of the easy task. Furthermore, the persistence of residual switch costs—switch costs that cannot be eliminated even with infinite preparation time—provides conclusive evidence that full guided activation requires the physical arrival of the sensory stimulus to complete the final competitive resolution at the motor output layer.

9.2 Proactive versus Reactive Cognitive Control

To capture the temporal heterogeneity of executive function within the Guided Activation framework, Todd S. Braver formulated the Dual Mechanisms of Control (DMC) framework. Braver proposed that the guided activation machinery can operate along a continuum defined by two fundamentally distinct computational modes: Proactive Control and Reactive Control.

Proactive Control represents an anticipatory, early-selection strategy. The prefrontal cortex actively retrieves and sustains goal representations well in advance of the anticipated stimulus or conflict event. Proactive control relies heavily on the sustained, tonic firing of the dorsolateral prefrontal cortex, continuously broadcasting top-down biases into posterior structures to configure the processing landscape beforehand. The advantages of proactive control are high behavioral accuracy, minimized reaction time costs, and the virtually complete elimination of distractibility. However, its primary drawback is high metabolic and cognitive expense: sustaining prefrontal attractors against decay demands continuous glucose consumption and exhaustive monoaminergic support.

Reactive Control, in contrast, is a “just-in-time,” late-correction strategy. In this operational mode, the prefrontal cortex remains largely quiescent or engaged in default processing until an ambiguous stimulus, a high-conflict event, or an explicit behavioral error actually occurs. At that critical moment, transient detection signals from the anterior cingulate cortex retroactively trigger a brief, emergency burst of prefrontal activation to resolve the immediate conflict. Reactive control is metabolically cheap, freeing up cognitive capacity for other demands, but it carries high performance penalties: slower reaction times, vulnerability to attentional capture, and elevated error rates on incongruent trials. The choice between proactive and reactive modes is continuously arbitrated by an organism’s assessment of environmental volatility, fatigue, and the Expected Value of Control.

9.3 Resolving Proactive and Retroactive Interference in Memory

The principles of top-down guided activation are not restricted to real-time sensory-to-motor transformations; they are equally essential for resolving competition within episodic and working memory retrieval. In memory tasks, individuals frequently confront intense interference from competing mnemonic traces:

• Proactive Interference: Previously learned information disrupts the encoding or retrieval of newly acquired information (e.g., trying to remember where you parked your car today, despite having parked in different spots on the previous twenty days).
• Retroactive Interference: Newly acquired information disrupts the retrieval of older, consolidated traces.

Neuroimaging and neuropsychological studies implicate the left ventrolateral prefrontal cortex (vlPFC, specifically Brodmann area 45/47 and the inferior frontal gyrus) as the critical hub for resolving mnemonic competition. According to guided activation principles, when a retrieval cue (such as “Where is the car?”) is presented, it automatically triggers feedforward activation across the medial temporal lobes and hippocampus, simultaneously activating all associated memory traces. The target trace and the competing traces enter a state of mutual inhibition within associative memory networks.

To prevent retrieval failure, the left vlPFC deploys a top-down contextual biasing signal directed into the medial temporal lobes. This biasing signal does not contain the detailed episodic memory itself; rather, it represents the specific spatiotemporal contextual criteria of the target (e.g., “today’s timestamp,” “visual features of the current parking lot”). This prefrontal contextual bias amplifies the target trace’s representation, allowing it to win the local competitive dynamic and breach the threshold for conscious recollection. Individuals with high working memory capacity possess superior prefrontal biasing efficiency, enabling them to resolve intense proactive interference rapidly, whereas individuals with prefrontal damage or diminished cognitive capacity exhibit profound source amnesia and catastrophic memory intrusions.

10. Comparison with Alternative and Complementary Theories of Prefrontal Function

10.1 Hierarchical Models of Prefrontal Organization

The original 2001 Guided Activation formulation treated the lateral prefrontal cortex largely as an integrated, uniform computational entity. However, subsequent anatomical and functional investigations revealed that the human prefrontal cortex is organized along distinct structural and functional gradients, giving rise to Hierarchical Models of Cognitive Control.

Spearheaded by Etienne Koechlin and further refined by David Badre and Mark D’Esposito, these models demonstrate a rostro-caudal gradient of abstraction spanning the frontal lobe:

• Sensory Control (Premotor Cortex): Selects motor actions based directly on immediate, univalent sensory signals (e.g., Green $\rightarrow$ Press Button).
• Contextual Control (Caudal dlPFC / Brodmann Area 8/9): Selects representations based on the immediate environmental context that determines how a sensory stimulus must be interpreted.
• Episodic Control (Rostrolateral dlPFC / Brodmann Area 46): Directs behavioral policies based on temporal context or historical events occurring across the broader behavioral session.
• Branching / Relational Control (Frontopolar Cortex / Brodmann Area 10): Occupies the absolute apex of the hierarchy, responsible for adjudicating between completely different behavioral goals, exploratory meta-policies, and long-term intentions.

Hierarchical control models do not contradict Guided Activation; rather, they provide its structural multi-tier extension. In an integrated hierarchical framework, guided activation operates iteratively down the rostro-caudal axis: frontopolar regions provide top-down modulatory biases that set the attractor states within mid-dorsolateral prefrontal regions, which in turn project biasing signals into premotor networks, which finally tilt the competitive balance among motor effectors in the primary motor cortex.

10.2 The Anterior Cingulate Cortex: Conflict Monitoring vs. Guided Activation

Any comprehensive model of cognitive control must define the precise division of labor between the dorsolateral prefrontal cortex (dlPFC) and the dorsal anterior cingulate cortex (dACC). While early imaging studies often lumped these structures together under an undifferentiated “frontoparietal control network,” Matthew Botvinick, Cameron Carter, Todd Braver, Deanna Barch, and Jonathan Cohen (2001) formulated the highly influential Conflict Monitoring Hypothesis, creating a formal mechanistic synthesis with Guided Activation.

The model explicitly assigns complementary roles to the two structures:

• The dACC as Conflict Monitor: The dACC continuously monitors information processing across downstream networks. It does not actively resolve conflict, nor does it maintain abstract task rules. Instead, it computes the level of mathematical energy or “cross-talk” resulting from the simultaneous activation of mutually incompatible response channels. High conflict occurs during incongruent Stroop trials or when an error is made. When conflict crosses a defined threshold, the dACC fires an alarm—a call for cognitive intervention.
• The dlPFC as Control Implementer: The lateral prefrontal cortex receives the conflict signal emanating from the dACC. In response, the dlPFC intensifies its top-down biasing signals, strengthening active rule representations and forcing downstream sensory-motor networks into compliance.

This dynamic feedback loop was mathematically captured by the Expected Value of Control (EVC) theory formulated by Amitai Shenhav, Matthew Botvinick, and Jonathan Cohen (2013). EVC theory posits that the dACC integrates multiple variables: the probability of conflict, the expected physical or cognitive effort, the risk of failure, and the prospective economic reward. From these variables, the dACC calculates the “expected value” of deploying top-down control and dictates to the dlPFC exactly how much metabolic effort and biasing gain ($\beta$) it should allocate to the current task. While alternative perspectives argue that the dACC is involved in foraging choices, error-likelihood estimation, or environmental surprise signaling, the functional coupling between dACC monitoring and dlPFC guided activation remains the dominant paradigm in computational neuroscience.

10.3 Free Energy Principle and Predictive Processing Approaches

In recent years, theoretical neuroscience has increasingly converged around the Free Energy Principle and Predictive Processing frameworks pioneered by Karl Friston. This perspective conceptualizes the brain as a hierarchical Bayesian inference machine whose overarching imperative is the minimization of surprise or variational free energy.

Within predictive processing architectures, the traditional view of feedforward sensory transmission is inverted: lower visual areas do not transmit raw sensory data up the hierarchy. Instead, descending feedback projections carry top-down prior predictions regarding the causes of sensory inputs, while ascending feedforward projections carry only the residual prediction errors—the variance left unexplained by the descending priors.

The Guided Activation Theory translates with extraordinary fidelity into predictive processing terminology:

• Prefrontal Goals as High-Level Priors: Prefrontal task demand units represent deep, abstract generative models of the environment, predicting which action-perception contingencies are valid within the current behavioral regime.
• Biasing as Precision Weighting: In Bayesian predictive coding, prediction errors must be dynamically scaled by their expected reliability or “precision” (inverse variance). The top-down modulatory bias of Miller and Cohen is computationally identical to precision weighting. When the prefrontal cortex biases an ambiguous sensory-motor pathway, it elevates the synaptic gain (precision) of the prediction errors generated along that specific channel, allowing them to drive belief updates and motor inference while down-weighting the precision of distractors.

Through this translation, guided activation is recognized not merely as an ad-hoc mechanism for suppressing habits, but as a direct biophysical implementation of hierarchical Bayesian inference under conditions of behavioral uncertainty.

11. Clinical Implications: Prefrontal Dysfunction in Neurological and Psychiatric Disorders

11.1 Schizophrenia and the Breakdown of Context Processing

Perhaps nowhere has the Guided Activation Theory provided more profound clinical insight than in the neurobiology of schizophrenia. While historically characterized by dramatic positive symptoms (hallucinations and delusions), the core, debilitating feature of schizophrenia is its profound cognitive deficit, particularly formal thought disorder, executive dysfunction, and catastrophic working memory failure.

Pioneering neuropsychological and computational work by Deanna Barch, Jonathan Cohen, and colleagues demonstrated that these cognitive impairments stem from a specific, primary deficit in the representation and maintenance of context within the prefrontal cortex. Using rigorous continuous performance paradigms such as the AX-CPT (where the letter “A” serves as a predictive context requiring a target response to “X”, while “B” indicates that “X” must be ignored), the authors demonstrated that patients with schizophrenia selectively fail trials requiring proactive context maintenance (e.g., BX trials). They do not suffer from an inability to respond to immediate stimuli; they fail to actively sustain the contextual “A” template required to suppress an automatic prepotent response when “X” appears.

This contextual failure maps directly onto microcircuit abnormalities revealed by post-mortem neuropathology:

• NMDA Receptor Hypofunction: Impairments in NR1/NR2A subunit transmission compromise the slow, recurrent excitation necessary to sustain prefrontal attractor states against decay.
• Parvalbumin Interneuron Pathology: Reductions in the synthesis enzyme GAD67 within parvalbumin-positive fast-spiking basket cells degrade the inhibitory microcircuitry, destroying the local lateral inhibition needed to clear obsolete representations and focus prefrontal gain.
• Gamma-Band Desynchronization: Electrophysiological recordings during cognitive control tasks confirm an inability of the schizophrenic prefrontal cortex to generate and sustain coherent 40 Hz gamma oscillations, resulting in a complete breakdown of top-down biasing signals directed into posterior processing streams.

11.2 Attention-Deficit/Hyperactivity Disorder (ADHD) and Control Deficits

Attention-Deficit/Hyperactivity Disorder (ADHD) provides another quintessential clinical manifestation of impaired guided activation. Long mischaracterized as a primary sensory-attentional deficit, modern cognitive neuroscience recognizes ADHD as an impairment in the regulatory machinery governing cognitive control, temporal foresight, and top-down biasing stability.

Neurodevelopmental structural MRI studies have demonstrated that children with ADHD exhibit a significant maturational delay (typically 2 to 3 years) in peak cortical thickness across the prefrontal cortex, accompanied by atypical white matter structural integrity along frontostriatal and frontoparietal tracts. Functionally, when individuals with ADHD engage in tasks requiring the suppression of prepotent responses (such as the Go/No-Go task or the Stop-Signal task), they display marked hypoactivation within the right inferior frontal gyrus and the anterior cingulate cortex. Their internal top-down biasing signals are transient, noisy, and incapable of providing continuous modulatory drive to posterior sensory-motor pathways. Consequently, their behavior is repeatedly captured by bottom-up perceptual salience, resulting in distractibility, motor impulsivity, and severe cognitive variability.

The pharmacological mechanisms of classical psychostimulants utilized to treat ADHD—such as methylphenidate and amphetamines—directly corroborate the Guided Activation framework. Psychostimulants block the dopamine transporter (DAT) and norepinephrine transporter (NET), selectively elevating extracellular catecholamine concentrations within the prefrontal cortex. By restoring prefrontal dopamine and norepinephrine levels to the optimal peak of their respective inverted-U dose-response curves, these agents stimulate post-synaptic $D_1$ and $\alpha_{2A}$ receptors. This neurochemical tuning enhances NMDA channel conductance, quiets spontaneous baseline firing, elevates the prefrontal signal-to-noise ratio, and directly restabilizes the top-down guided activation signals required to suppress distracting environmental noise.

11.3 Substance Use Disorders, Addiction, and Compulsive Behaviors

Substance Use Disorders (SUDs) and behavioral addictions represent a catastrophic functional uncoupling between prefrontal top-down control systems and subcortical habit-formation networks. As recreational drug use transitions into chronic, compulsive chemical dependency, the brain undergoes severe neuroplastic reorganizations across its decision-making architecture.

Structural neuroimaging in chronic addiction reveals significant gray matter volumetric reductions across the dorsolateral prefrontal cortex, ventromedial prefrontal cortex, and insular mantle. Simultaneously, neurochemical analyses confirm severe down-regulations of striatal and prefrontal dopamine $D_2$ receptor densities. Within the Guided Activation framework, this pathology manifests as a devastating double hit:

• Hyperactive Subcortical Drive: Sensitized drug-cue-associated assemblies within the ventral striatum and extended amygdala deliver massive, bottom-up motivational surges whenever drug-related cues are encountered.
• Prefrontal Hypofrontality: Structural and functional atrophy of the prefrontal cortex abolishes its capacity to maintain alternative, goal-directed value representations or inject top-down modulatory biases to counteract the craving.

The railway switches are permanently jammed in the direction of the habit-driven subcortical track. Even when individuals express genuine, conscious desire to achieve sobriety, the presentation of a conditioned drug cue bypasses the degraded prefrontal control machinery, driving compulsory drug-seeking locomotion. Contemporary clinical interventions directly target this guided activation deficit: repetitive Transcranial Magnetic Stimulation (rTMS) protocols applied over the left dlPFC seek to artificially enhance cortical excitability, re-establishing sufficient top-down modulatory bias to restore impulse control and suppress cue-induced cravings.

12. Modern Revisions, Current Debates, and Future Directions in Guided Activation

12.1 Dynamic Coding, Travelling Waves, and Neural Geometry

The classic 2001 formulation of Guided Activation was conceptually married to the classical electrophysiological view of persistent, static delay-period firing. In that traditional model, single prefrontal neurons were presumed to fire continuously at an elevated, stationary rate to maintain a rule across time. Over the past decade, however, the paradigm of prefrontal electrophysiology has shifted radically toward Dynamic Population Coding and High-Dimensional Neural Manifolds.

Multi-channel recording technologies (such as Neuropixels probes) have revealed that when individual prefrontal neurons are monitored across a delay period, their firing rates are rarely static. Instead, single units display complex, heterogeneous, and time-varying firing patterns—rising, falling, bursting, and falling silent across the delay interval. Yet, when analyzed at the level of the entire neural population using dimensionality-reduction techniques (such as Principal Component Analysis, Tensor Component Analysis, or dPCA), these seemingly chaotic individual dynamics trace out remarkably stable, predictable trajectories along low-dimensional curved surfaces known as neural manifolds.

Furthermore, cutting-edge work by Earl K. Miller, Mikael Lundqvist, and colleagues has shown that top-down prefrontal communication is organized across space and time through travelling waves. Beta and gamma oscillations do not simply pulse in uniform synchrony across the cortex; they propagate across the prefrontal surface in spatial waves traveling from dorsal to ventral regions. These travelling waves serve as a dynamic clocking mechanism, sequentially gating and sorting the readouts of different neural ensembles. Far from undermining the Guided Activation Theory, dynamic coding and travelling waves enrich it: they reveal that prefrontal top-down biasing is not a rigid, static clamp, but a dynamic, time-varying trajectory that guides processing along flexible, multidimensional contours.

12.2 Artificial Intelligence, Deep Learning, and Cognitive Control Architectures

The conceptual framework of Guided Activation has exerted a massive, transformative influence on modern Artificial Intelligence (AI), machine learning, and deep neural network design. As deep learning models exploded in complexity during the 2010s, artificial intelligence engineers encountered the exact computational constraints that biological nervous systems solved millions of years ago: the challenges of catastrophic forgetting, interference between competing tasks, and the inability to generalize out-of-distribution.

The revolutionary breakthrough of the Transformer Architecture (Vaswani et al., 2017) and its self-attention and cross-attention mechanisms represents a computational analog of Guided Activation:

• Queries, Keys, and Values: In a transformer network, incoming data tokens are mapped into “Key” and “Value” representations (equivalent to posterior sensory-motor pathways). “Query” vectors dynamically interrogate the keys to compute an attention weight matrix.
• Attention as Top-Down Biasing: This attention matrix selectively multiplies the values, exponentially amplifying task-relevant information while zeroing out irrelevant features. Just as in Miller and Cohen’s model, the core computations are not completely recomputed; instead, dynamic, contextual vectors continuously scale the flow of information through intermediate pathways.

Similarly, state-of-the-art developments in Meta-Reinforcement Learning demonstrate that recurrent neural networks (RNNs) trained on diverse distributions of tasks spontaneously evolve internal computational dynamics mirroring the prefrontal-basal ganglia control loop. The network’s recurrent internal activations spontaneously organize into attractor manifolds that hold task contexts and project top-down modulatory weights into feedforward layers, proving that guided activation principles are mathematically optimal for any resource-constrained architecture navigating complex, multi-task environments.

12.3 Open Questions and the Next Frontier in Cognitive Control

Despite more than two decades of extraordinary empirical and computational triumph, the Guided Activation Theory stands before several critical open questions and theoretical frontiers:

The Meta-Control Dilemma: If the prefrontal cortex configures posterior pathways via top-down biasing, what system configures the prefrontal cortex? How does the brain decide, autonomously and dynamically, what the current goal ought to be without falling into the infinite regress of an internal homunculus? Emerging models suggest that meta-control is an emergent property of decentralized loops connecting the frontomedial cortex, the hippocampus (providing episodic context), and the ascending monoaminergic systems evaluating uncertainty, survival imperatives, and affective drive.

Metabolic and Energetic Constraints: What is the biological currency of cognitive control? Why does sustained prefrontal biasing feel subjectively effortful, and why does prolonged executive engagement lead to mental exhaustion? Contemporary research is exploring metabolic constraints, including the buildup of extracellular glutamate within prefrontal synapses during high-load processing (which degrades the signal-to-noise ratio and induces fatigue) and the finite availability of astrocytic glycogen reserves.

The Role of Non-Cortical Hubs (Mediodorsal Thalamus and Claustrum): While classical guided activation focused primarily on direct cortico-cortical connections, contemporary systems neuroscience has revealed that subcortical hubs play an indispensable coordinating role. The mediodorsal (MD) nucleus of the thalamus maintains reciprocal connections with the dlPFC and acts as an active, sustained amplifier of prefrontal attractor dynamics. The claustrum—a thin, hyper-connected subcortical sheet—is uniquely poised to orchestrate widespread synchronous oscillations across the cortical mantle. The next generation of Guided Activation models will transition from a “prefrontal-centric” framework to a fully distributed, multi-regional “corticothalamocortical” circuit architecture.

Conclusion: The Enduring Legacy of Guided Activation

When Earl K. Miller and Jonathan D. Cohen published their integrative theory of prefrontal cortex function in 2001, they achieved a paradigm shift in cognitive neuroscience. By conceptualizing the prefrontal cortex not as a mysterious, all-knowing command center, but as a mechanistic source of top-down modulatory bias, they dismantled the homuncular ghost that had haunted executive function research for over a century.

The Guided Activation Theory revealed that the human capacity for adaptive, voluntary, and flexible behavior does not require a conscious agent pulling levers inside the frontal lobes. Instead, cognitive control is the natural, emergent consequence of a competitive, distributed neural architecture:

• Hardwired sensory-motor pathways provide speed, automaticity, and efficiency for routine survival behaviors.
• The prefrontal cortex, with its dense reciprocal connections, slow biophysical kinetics, and capacity for abstract categorical representation, provides the top-down biasing counterweight.
• By subtly adjusting the internal conductance of the nervous system, prefrontal signals ensure that our internal goals, rather than the immediate physical salience of the environment, dictate our actions, our thoughts, and our destinies.

From the single-unit electrophysiology of the principal sulcus to the dynamic functional networks revealed by human neuroimaging; from the biophysical realities of NMDA and dopamine channels to the mathematical architecture of modern artificial intelligence; and from the clinical management of schizophrenia and ADHD to the theoretical frontiers of predictive processing and neural manifolds, the Guided Activation Theory remains one of the most durable, intellectually generative, and profoundly elegant syntheses in the history of brain science.

References

• Aston-Jones, G., & Cohen, J. D. (2005). An integrative theory of locus coeruleus-norepinephrine function: Adaptive gain and optimal performance. Annual Review of Neuroscience, 28, 403-450. https://doi.org/10.1146/annurev.neuro.28.061604.135709
• Baddeley, A. D., & Hitch, G. (1974). Working memory. In G. H. Bower (Ed.), The Psychology of Learning and Motivation (Vol. 8, pp. 47-89). Academic Press. https://doi.org/10.1016/S0079-7421(08)60452-1
• Badre, D., & D’Esposito, M. (2009). Is the rostro-caudal axis of the frontal lobe hierarchical? Nature Reviews Neuroscience, 10(9), 659-669. https://doi.org/10.1038/nrn2607
• Barch, D. M., Carter, C. S., Braver, T. S., Sabb, F. W., MacDonald, A., Noll, D. C., & Cohen, J. D. (2001). Selective deficits in the maintenance of context in schizophrenia: Functional magnetic resonance imaging of an AX-CPT task. Journal of Abnormal Psychology, 110(2), 240-258. https://doi.org/10.1037/0021-843X.110.2.240
• Botvinick, M. M., Carter, C. S., Braver, T. S., Barch, D. M., & Cohen, J. D. (2001). Conflict monitoring and cognitive control. Psychological Review, 108(3), 624-652. https://doi.org/10.1037/0033-295X.108.3.624
• Braver, T. S. (2012). The variable nature of cognitive control: A dual mechanisms framework. Trends in Cognitive Sciences, 16(2), 106-113. https://doi.org/10.1016/j.tics.2011.12.010
• Buschman, T. J., & Miller, E. K. (2007). Top-down versus bottom-up control of attention in the prefrontal and posterior parietal cortices. Science, 315(5820), 1860-1862. https://doi.org/10.1126/science.1138071
• Cohen, J. D., Dunbar, K., & McClelland, J. L. (1990). On the control of automatic processes: A parallel distributed processing account of the Stroop effect. Psychological Review, 97(3), 332-361. https://doi.org/10.1037/0033-295X.97.3.332
• Desimone, R., & Duncan, J. (1995). Neural mechanisms of selective visual attention. Annual Review of Neuroscience, 18, 193-222. https://doi.org/10.1146/annurev.ne.18.030195.001205
• Duncan, J. (2001). An adaptive coding model of neural function in prefrontal cortex. Nature Reviews Neuroscience, 2(11), 820-829. https://doi.org/10.1038/35097575
• Frank, M. J., Loughry, B., & O’Reilly, R. C. (2001). Interactions between frontal cortex and basal ganglia in working memory: A computational model. Cognitive, Affective, & Behavioral Neuroscience, 1(2), 137-160. https://doi.org/10.3758/CABN.1.2.137
• Freedman, D. J., Riesenhuber, M., Poggio, T., & Miller, E. K. (2001). Categorical representation of visual stimuli in the primate prefrontal cortex. Science, 291(5502), 312-316. https://doi.org/10.1126/science.291.5502.312
• Fries, P. (2005). A mechanism for cognitive dynamics: Neuronal communication through coherence. Trends in Cognitive Sciences, 9(10), 474-480. https://doi.org/10.1016/j.tics.2005.08.011
• Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127-138. https://doi.org/10.1038/nrn2787
• Goldman-Rakic, P. S. (1995). Cellular basis of working memory. Neuron, 14(3), 477-485. https://doi.org/10.1016/0896-6273(95)90304-6
• Koechlin, E., Ody, C., & Kouneiher, F. (2003). An architecture of cognitive control in the human prefrontal cortex. Science, 302(5648), 1181-1185. https://doi.org/10.1126/science.1088595
• Lundqvist, M., Herman, P., & Miller, E. K. (2018). Working memory: Delay activity, oscillations, and breakthroughs to new models. Current Opinion in Neurobiology, 52, 70-81. https://doi.org/10.1016/j.conb.2018.04.017
• Miller, E. K., & Cohen, J. D. (2001). An integrative theory of prefrontal cortex function. Annual Review of Neuroscience, 24, 167-202. https://doi.org/10.1146/annurev.neuro.24.1.167
• Moore, T., & Armstrong, K. M. (2003). Selective gating of visual signals by microstimulation of frontal cortex. Nature, 421(6921), 370-373. https://doi.org/10.1038/nature01341
• Norman, D. A., & Shallice, T. (1986). Attention to action: Willed and automatic control of behavior. In R. J. Davidson, G. E. Schwartz, & D. Shapiro (Eds.), Consciousness and Self-Regulation (pp. 1-18). Springer US. https://doi.org/10.1007/978-1-4757-0629-1_1
• Rigotti, M., Barak, O., Warden, M. R., Wang, X.-J., Daw, N. D., Miller, E. K., & Fusi, S. (2013). The importance of mixed selectivity in complex cognitive tasks. Nature, 497(7451), 585-590. https://doi.org/10.1038/nature12160
• Shenhav, A., Botvinick, M. M., & Cohen, J. D. (2013). The expected value of control: An integrative theory of anterior cingulate cortex function. Neuron, 79(2), 217-240. https://doi.org/10.1016/j.neuron.2013.07.007
• Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30, 5998-6008. https://arxiv.org/abs/1706.03762
• Wang, X.-J. (2002). Probabilistic decision making by slow reverberation in recurrent networks. Neuron, 36(5), 955-968. https://doi.org/10.1016/S0896-6273(02)01092-9