The construction of the Tower of Béla - Adventures in Anaglyphs - Part One - By Daniel Scott Matthews

 The Layered Predictive Processing Model of Consciousness

Executive Summary

This document provides a comprehensive analysis of the "Layered Predictive Processing Model," a theoretical framework that seeks to unify neuroscience, computational modeling, and evolutionary psychology to explain the architecture of human consciousness. The model posits that the mind is built upon a foundational predictive engine, with higher-order cognitive functions like verbal thought and visuospatial reasoning operating as distinct, computationally specialized layers.

The most critical takeaways are as follows:

• A Unified Brain Architecture: The model's foundation rests on two strongly supported neuroscientific principles:
  1. Uniform Hardware: The neocortex is composed of millions of uniform, repurposable computational units known as "cortical columns." This modular design enabled the rapid evolutionary expansion of the human brain.
  2. Universal Algorithm: These columns run a universal algorithm of "predictive processing," where the brain's primary function is to constantly generate a model of the world to predict sensory input and minimize the error between prediction and reality.
• Dual High-Level Simulation Engines: Built upon this predictive foundation are at least two distinct cognitive systems for advanced reasoning:
  1. The Verbal Layer: Inner speech functions as a biological analogue to a Large Language Model (LLM). It is a generative, predictive system that manipulates symbolic information. However, unlike current LLMs, it is deeply embodied—grounded in the simulated action of speech—and is wielded by executive functions as a tool for thought.
  2. The Visuospatial Layer: Mental imagery functions as a non-verbal, analog simulation engine, akin to a physics or CAD program. It operates by re-purposing the brain's own perceptual hardware to manipulate holistic representations according to quasi-physical constraints, as demonstrated by mental rotation tasks.
• Evolutionary Mandate and Sentient Emergence: The entire cognitive architecture is framed as a product of evolution, designed to enhance survival and reproduction by predicting threats and opportunities ("find food, avoid death"). The model's most profound claim is that sentience itself is an emergent property of this complex predictive system. In this view, subjective experience is not an external force "hacking" the hardware for novel purposes; rather, it is the system's own high-level experience of its global predictive state. Affective states like joy or fear are the ultimate feedback signals, representing the successful or failed minimization of prediction error, which in turn guides all behavior.

The Foundational Architecture: A Predictive Neocortex

The model's viability rests on a powerful synthesis of the brain's physical structure (hardware) and its core computational principle (software).

The Canonical Microcircuit: Uniform, Repurposable Hardware

The human neocortex, despite its diverse functions, exhibits a profound structural uniformity.

• The Columnar Hypothesis: The cortex is composed of an estimated two to four million repeating, structurally similar subunits called "cortical columns." These vertical columns, extending through the six layers of the neocortex, are considered the fundamental modular units of computation.
• Functional Interchangeability: The function of a cortical column is determined not by its internal structure but by its inputs and outputs. Landmark experiments demonstrated this principle by rerouting visual inputs to the auditory cortex in ferrets, causing the auditory cortex to begin processing information visually.
• Evolutionary Scalability: The existence of a uniform, repurposable blueprint (the cortical column) is the most plausible mechanism for the human neocortex's rapid threefold expansion over the last three million years. Natural selection could scale intelligence by massively duplicating this existing, versatile design.

The Predictive Processing Framework: A Universal Algorithm

The universal hardware of the cortical columns is proposed to run a universal operating system known as the Predictive Processing (PP) framework.

• The Brain as a Prediction Engine: The PP framework posits that the brain is not a passive reactor to sensory input but an active, top-down prediction engine. Its fundamental goal is to minimize "prediction error"—the mismatch between its expected sensory input and the actual data it receives.
• Hierarchical Bayesian Inference: The brain's internal world model is organized hierarchically. Higher levels generate predictions about the activity of lower levels. If a prediction is accurate, the signal is cancelled out. If there is a mismatch, an "error signal" is propagated up the hierarchy to update the model. This is an extremely efficient scheme that prioritizes novel information.
• Hardware-Software Integration: The six-layered anatomical structure of the cortical column appears to be a near-perfect physical implementation of the PP algorithm. Top-down predictions are carried by feedback connections from deep cortical layers, while bottom-up error signals are carried by feedforward connections targeting superficial layers. The "generic recogniser" is more precisely a "generic predictor."

High-Level Cognitive Systems: Dual Simulation Engines

The model proposes that this predictive foundation supports distinct, high-level simulation engines for different types of thought.

The Verbal Layer: Inner Speech as a Biological LLM

The inner monologue is modeled as a cognitive layer analogous to a Large Language Model (LLM).

• Neurobiological Basis: Inner speech is not an abstract process but a simulated action. Neuroimaging and Brain-Computer Interface (BCI) studies show that inner speech activates the same motor-planning areas of the brain used for overt speech (e.g., Broca's area, supplementary motor area), but with the final muscular output inhibited. This grounds verbal thought in the sensorimotor system.
• Convergent Function: Both human language centers and LLMs operate on a predictive principle (next-word prediction). Research has demonstrated a direct linear alignment between the internal representations of an LLM and the neural activity patterns in the human brain during language processing, providing strong evidence for a shared computational framework.
• Powerful but Incomplete Analogy: While the core function is similar, key differences exist. LLMs are stateless predictors trained on disembodied text, lacking intrinsic goals or agency. Human inner speech is a tool used by a broader cognitive system for planning, self-regulation, and belief updating. Therefore, the LLM is a model of the generative language faculty, not the complete agent that uses it to think.

The following table summarizes the key distinctions:

Attribute	Human Inner Speech	Large Language Model (LLM)
Underlying Architecture	Biological neural networks (cortical columns) within a massively parallel, hierarchical brain.	Artificial neural networks (Transformers), typically run on sequential processors.
Core Function	Next-word prediction as one function within a broader cognitive architecture dedicated to minimizing global prediction error for survival.	Next-token prediction as the primary, and often sole, training objective.
Training Data	Embodied, multimodal, real-time, lived experience; grounded in sensory input, motor actions, and social interaction.	Disembodied, static, massive text and code corpora; lacks grounding in the physical world.
Generative Process	A simulation of the physical act of speech, heavily influenced by internal goals, emotions, and a world model.	Statistical sampling from a learned probability distribution over a vocabulary of tokens.
Embodiment/Grounding	Deeply grounded in the brain's sensory, motor, and affective systems; words have sensorimotor meaning.	Ungrounded symbolic manipulation; meaning is derived purely from statistical co-occurrence with other symbols.
Agency & Goals	Serves the intrinsic goals of an integrated, autonomous agent (e.g., planning, self-regulation, problem-solving).	Lacks intrinsic agency or goals; behavior is determined by the user's prompt and the system's alignment training.

The Visuospatial Layer: Mental Imagery as an Analog Simulator

Complementing the symbolic verbal layer is a non-verbal system for visuospatial reasoning.

• Shared Neural Substrate: The "mind's eye" is not a metaphor. Neuropsychological and neuroimaging evidence confirms that generating mental images co-opts the same neural hardware used for actual visual perception, including early areas like the primary visual cortex (V1). Patients with perceptual deficits (e.g., hemispatial neglect) exhibit corresponding deficits in their mental imagery.
• Analog Simulation: This layer operates on principles that are continuous and analog, not discrete and symbolic. The classic mental rotation experiments by Shepard and Metzler demonstrated this: reaction time to identify rotated objects is a linear function of the angle of rotation, suggesting the brain performs a continuous, simulated rotation that mimics physical reality.
• Complementary Computational Paradigm: The verbal and visuospatial layers represent two fundamentally different computational styles—one symbolic and statistical, the other analog and geometric. The power of human intelligence derives from the seamless integration of these two complementary engines, allowing for complex problem-solving that synthesizes logical reasoning with intuitive, spatial simulation.

Evolutionary Context and Emergence of Sentience

The model places this cognitive architecture within a broader evolutionary and philosophical framework.

The Evolutionary Mandate: "Find Food, Avoid Death"

From an evolutionary psychology perspective, the entire predictive architecture is an adaptation for survival and reproduction.

• Simulation as a Survival Tool: The ability to run internal simulations of potential scenarios—whether verbal or visuospatial—is a powerful adaptive tool, allowing for planning and risk assessment without incurring real-world costs.
• Mismatch Theory: The human brain's core programs were shaped for the ancestral Environment of Evolutionary Adaptedness (EEA). Many modern behaviors, both rational and irrational, can be understood as these ancient survival programs running on novel, modern inputs. For example, cravings for calorie-dense food, once adaptive, are now maladaptive in environments of abundance.
• Culture as Repurposed Survival: Even the highest forms of culture, such as art and science, can be interpreted as complex repurposings of fundamental survival drives, such as status signaling (to attract mates) or foraging (the drive for discovery).

Sentience as an Emergent Property

The model's final claim—that "sentience hacks the hardware"—is reframed through the scientific lens of emergence.

• Neurobiological Emergentism: Sentience is proposed to be a "weakly" emergent property that arises naturally and lawfully from the complex, hierarchical, and densely interconnected interactions of the brain's predictive systems. It is a novel, system-level property not reducible to its individual neural components.
• Convergence with Consciousness Theories: The proposed architecture is compatible with other major theories of consciousness. It provides a plausible cognitive structure that could implement the information "broadcast" of Global Workspace Theory (GWT) or possess the high degree of integrated information required by Integrated Information Theory (IIT).
• Sentience as the Ultimate Predictive Signal: The most elegant synthesis is to view sentience not as a separate entity but as the subjective experience of the system's own global predictive state.
  • Positive Affect (e.g., joy, curiosity) is the experience of low or decreasing global prediction error, signaling that the organism's world model is accurate and effective.
  • Negative Affect (e.g., pain, anxiety) is the experience of high or catastrophic prediction error, serving as a powerful alarm that compels the organism to update its model or change its environment.

In this integrated view, the "hack" is the system's discovery and mastery of its own operating principles, learning to manipulate its predictive machinery to generate states experienced as meaningful, beautiful, or profound.

Conclusion and Future Directions

The Layered Predictive Processing Model is a scientifically grounded and highly coherent framework that successfully synthesizes the brain's hardware, software, high-level functions, and evolutionary history.

Key Refinements:

1. From Layers to Integrated Systems: The model's power can be enhanced by moving from a concept of distinct "layers" to one of deeply integrated, computationally distinct systems whose constant interplay gives rise to complex thought.
2. From "Hacking" to Integrated Feedback: Sentience should be viewed not as an external agent "hacking" the system, but as the emergent, subjective experience of the system's global predictive state—the ultimate feedback and control signal for the entire organism.

Key Questions for Future Research:

• System Integration: How does the brain arbitrate between predictions from the verbal and visuospatial systems and bind them into a unified experience?
• Neurobiological Mapping: Can the real-time flow of predictions and error signals between these systems be mapped during complex problem-solving?
• Artificial Intelligence: Could a new AI architecture combining a symbolic LLM with an analog physics engine, all governed by a global error-minimization objective, lead to more flexible and general intelligence?