Cognitive Reality Monitoring Network (CRMN)

• CRMN is a distributed executive architecture in the lateral PFC that gates paired generative and inverse models to enable metacognition and conscious access.
• It integrates multisensory, motor, and abstract cognitive inputs through responsibility-weighted signals that drive adaptive learning and action.
• Empirical evidence links CRMN’s dynamic gating mechanisms to rapid learning, improved metacognitive confidence, and accurate agency perception in neural studies.

The Cognitive Reality Monitoring Network (CRMN) is a distributed executive architecture posited to reside in the lateral prefrontal cortex (PFC), orchestrating the dynamic selection and gating of internal model pairs for perception, action, and learning. Anchored in computational neuroscience and neuroimaging evidence, CRMN implements the algorithmic substrate for metacognition and consciousness by monitoring, gating, and updating a modular hierarchical arrangement of generative-inverse model pairs, with explicit responsibility signals driving selection and plasticity across parallel sensory, motor, and higher cognitive domains (Kawato et al., 2021).

1. Architectural Foundations

CRMN is configured as a gating and monitoring hub in the lateral PFC, integrating inputs from all cortical modules (spanning sensory to motor streams, as well as abstraction levels) and basal ganglia. Each module within the system is defined by a conjugate pair: a generative (forward) model $g_{i,k}$, mapping high-level states downward to predict sensory or motor signals, and an inverse (recognition/control) model $f_{i,k}$, mapping observations or goals upward to infer latent states or motor commands. Modules are topologically organized in a two-dimensional grid: parallel modality/task context axis $(i=1, \dots, N)$ and hierarchical axis $(k=1, \dots, K)$ representing increasing abstraction.

Information processing within each module proceeds via recurrent loops: sensory or motor signals traverse through the inverse model, reconstructed by the forward model, and compared to the original input, generating a mismatch signal $m_{i,k}$. Simultaneously, each module-partnered representation is connected to the basal ganglia for reward prediction, yielding a reward-prediction error $\delta_{i,k}$. The CRMN continuously aggregates mismatch and reward-error signals from all modules, computes a normalized "responsibility" signal $\rho_{i,k}$, and feeds this back across the grid to gate perception, action, and learning updates.

2. Core Computational Principles

The key operational signals and mechanisms in CRMN are formalized as follows (suppressing the hierarchy index $k$ for notational clarity):

• Generative Model: $\hat x_i = g_i(z_i; \theta^g_i)$, predicting a lower-level sensory/motor pattern from a latent state.
• Inverse Model: $\hat z_i = f_i(x_i; \theta^f_i)$, inferring a latent or command state from observed input.
• Mismatch Signal: $f_{i,k}$0 with $f_{i,k}$1, quantifying reconstruction consistency.
• Reward Prediction Error: $f_{i,k}$2, measuring the local discrepancy between received reward $f_{i,k}$3 and predicted value $f_{i,k}$4.
• Responsibility Signal: An un-normalized drive $f_{i,k}$5 (with scalar weights $f_{i,k}$6) is transformed by a softmax at inverse temperature $f_{i,k}$7:

$f_{i,k}$8

The entropy $f_{i,k}$9 serves as a measure of the network's certainty in its selection.

• Gating: Perceptual and action relevant signals, as well as parameter updates for learning, are all weighted by the corresponding $(i=1, \dots, N)$0.

3. Gating and Learning Dynamics

CRMN governs both the selection and plasticity of internal model modules via responsibility-weighted gating:

• Perceptual Gating: In sensory streams, perceived latent states are formed as convex combinations of inverse-model outputs $(i=1, \dots, N)$1.
• Motor Gating: In motor streams, actions are similarly derived from responsibility-weighted controller outputs.
• Learning Updates: Parameter updates for forward and inverse models, and value functions, are all scaled by $(i=1, \dots, N)$2:
  • Forward model: $(i=1, \dots, N)$3
  • Inverse model: $(i=1, \dots, N)$4
  • Value function: $(i=1, \dots, N)$5
  • This selection ensures that only modules with high responsibility are updated, effectively concentrating representational and policy learning in the most reliable circuits.

Pseudocode for the full gating and learning cycle is provided in the primary source (Kawato et al., 2021).

4. Consciousness and Metacognition Mechanisms

CRMN operationalizes metacognition as the ability to "read out" internal latent states and their computations from the modules with the highest $(i=1, \dots, N)$6. Conscious content is determined by the sharpness of the responsibility distribution:

• When one $(i=1, \dots, N)$7, entropy $(i=1, \dots, N)$8 is near zero, and the corresponding module's content is globally broadcast in the PFC, resulting in full conscious awareness.
• When $(i=1, \dots, N)$9 is flat (high entropy), no single module dominates, and the system lacks clear conscious content.
• A threshold $(k=1, \dots, K)$0 can be posited, below which a conscious “workspace” is established.

Metacognitive judgments, confidence reporting, and executive readout are grounded in this responsibility-based selection.

5. Empirical and Physiological Correlates

CRMN theory bridges computational models with a range of neurophysiological and behavioral findings:

• Cerebellar Internal Models: Foundational work (Kawato et al. 1987; Shidara et al. 1993) showed that cerebellar circuits implement forward/inverse model pairs for motor control, extended here to the cognitive domain via PFC-CRMN.
• Tool-Learning fMRI: PFC and cerebellar circuits flexibly and selectively gate internal models during tool learning (Imamizu et al. 2000, 2003), with CRMN providing the underlying module-selection signal.
• Reinforcement Learning (Decoded-Neurofeedback): Experiments (Cortese et al. 2020) revealed that, during RL, widespread cortical activity quickly focuses into PFC-basal ganglia loops, closely matching the responsibility-driven module selection mechanisms predicted by CRMN. Participants with higher metacognitive confidence accuracy in PFC exhibited faster RL and smaller basal ganglia decoded reward-prediction errors, consistent with a negative coupling between $(k=1, \dots, K)$1 (responsibility/confidence) and $(k=1, \dots, K)$2.
• Awareness Paradox: Behaviorally effective but unconscious neural patterns (e.g., in V1) manifest as large local mismatches $(k=1, \dots, K)$3, small $(k=1, \dots, K)$4, and high gating entropy, resulting in no conscious broadcast (Shibata et al. 2011; Cortese et al. 2016).
• Sense-of-Agency and Deficits: CRMN explains agency illusions and their breakdown in schizophrenia by mismatch between top-down predictions and bottom-up feedback (see Blakemore et al. 2000). In phenomena such as aphantasia and blindsight, weak sensory or top-down signals produce large $(k=1, \dots, K)$5, low $(k=1, \dots, K)$6, and thus no conscious content, despite preserved residual processing under uniform $(k=1, \dots, K)$7 (Kawato et al., 2021).

6. Functional Implications and Theoretical Integration

CRMN transforms modular, hierarchical internal models into a system with three distinct properties:

1. Metacognition: The ability to access and monitor the computations and latent states of the current "most responsible" module.
2. Consciousness: Linked to the selectivity (low entropy) of responsibility-weighted gating; only when CRMN confidently selects a module does its content become available to global broadcasting in PFC.
3. Sample-Efficient Learning: By dynamically restricting learning and perceptual/action involvement to modules with highest empirical accuracy (low $(k=1, \dots, K)$8) and reward-predictive value (low $(k=1, \dots, K)$9), CRMN enables efficient credit assignment and rapid learning in large, high-dimensional model spaces (Kawato et al., 2021).

A plausible implication is that CRMN provides a scalable, neurobiologically plausible mechanism for integrating internal models, metacognitive monitoring, conscious access, and reinforcement-based learning within unified cortical architectures.