Delusions Emerge from Generative Model Reorganisation rather than Faulty Inference: Insights from Hybrid Predictive Coding

This paper proposes a hybrid predictive coding model demonstrating that delusions arise not from faulty inference but as an adaptive reorganisation of generative models triggered by inadequate initialisation and excessive certainty, thereby explaining their thematic stability and resistance to contradictory evidence.

The Gist

The Big Idea: It's Not a Broken Calculator, It's a Rewritten Map

Imagine your brain is a super-smart detective trying to solve a mystery every time you see something new. This detective uses a "best guess" strategy called Predictive Coding.

Usually, the detective works like this:

1. The Guess: You see a blurry shape. Your brain quickly guesses, "That's a cat!" (This is the amortised guess).
2. The Check: The detective looks closer. "Wait, it has a tail and whiskers. Yes, it's a cat." (This is the inference or checking process).
3. The Update: If the guess was wrong, the detective updates their mental map of the world so they don't make the same mistake next time.

The Problem: In people with delusions (like in schizophrenia), the detective seems to get stuck. They see a harmless knife in a kitchen, but they are convinced it's a weapon. Even when you show them the knife is just for cooking, they refuse to change their mind.

The Old Theory: Scientists used to think the detective was just "bad at math." They thought the brain was weighing the evidence wrong—maybe listening too much to the "knife" and not enough to the "kitchen."

The New Theory (This Paper): The authors, Victor Navarro and Christoph Teufel, say the detective isn't bad at math. Instead, the map itself has been rewritten. The brain has reorganized its entire understanding of reality to make the delusion make sense.

---

The Experiment: The "Digit vs. Letter" Mix-Up

To prove this, the researchers built a computer brain (a neural network) and taught it to recognize handwritten numbers (0–5) and letters (A, B, C, etc.).

They set up a "Progenitor" model (the healthy brain) that learned perfectly. Then, they created two "Twins":

1. The Control Twin: Continued learning normally.
2. The Delusional Twin: The researchers swapped the brain's shortcuts.

The Analogy: Imagine you have two dictionaries. One is for Numbers, and one is for Letters.

• Normally, when you see a "4", you open the Number Dictionary.
• In the "Delusional Twin," the researchers forced the brain to open the Letter Dictionary when it saw a "4".

So, the brain's first guess (the shortcut) for the number "4" was "A".

What Happened Next? The "Reorganization"

Here is where it gets fascinating. The researchers didn't just leave the brain confused; they made the brain stubborn. They told the brain, "You are 100% sure that '4' is an 'A'. Do not change your mind."

1. The Struggle: At first, the brain tried to fix the mistake. It saw the shape of a "4" and thought, "But this looks like a 4!" It tried to correct the "A" guess.
2. The Pivot: But because the brain was forced to be so certain about being an "A," it couldn't just say "I was wrong." Instead, it started rewriting its internal map.
3. The Result: The brain changed its understanding of what a "4" looks like. It reorganized its neural connections so that the shape of a "4" actually did look like an "A" in its new reality.

The Key Finding:
Once the map was rewritten, the brain wasn't confused anymore. It wasn't making "mistakes."

• Prediction Error: In a normal brain, seeing a "4" and thinking "A" creates a huge "error signal" (a feeling of wrongness).
• In the Delusional Brain: Because the map was rewritten, the "A" guess fit the "4" shape perfectly. The error signal disappeared. The brain felt completely certain and correct.

Why This Matters: The "Ontological Shift"

This explains the two biggest mysteries of delusions:

1. Why are they so stubborn?
You can't argue with someone using logic because, in their new reality, the logic is perfect. If you tell them, "That's a knife, not a weapon," they hear, "That's a weapon, not a knife." Their brain has reorganized so that the "weapon" explanation is the one that fits the sensory data best. They aren't ignoring evidence; they are interpreting it through a completely different map.

2. Why do they have themes?
The paper suggests that delusions aren't random. They are the result of the brain trying to make sense of a specific "glitch" (like the wrong dictionary being opened). The brain builds a consistent story around that glitch.

The Takeaway for Treatment

The authors suggest that current treatments often try to argue with the "detective" (the inference). They try to say, "Look, that's not a weapon!"

But this paper suggests we need to retrain the map (the generative model).

• Early Intervention: If you catch the delusion early, the map hasn't been rewritten yet. You just need to stop the brain from getting stuck in the wrong "dictionary."
• Late Intervention: If the delusion has been there for years, the map is deeply rewritten. You can't just argue; you have to help the patient slowly "unlearn" the new map and relearn the old, healthy one.

Summary Metaphor

Imagine you are driving in a city.

• Normal Brain: You see a red light, you stop. If you see a green light, you go.
• Faulty Inference Theory: You see a green light, but you think "Red" because your eyes are broken.
• This Paper's Theory: Your eyes work fine. But someone rewrote the city's traffic laws. In this new city, Green means Stop. You stop at the green light. You aren't crazy; you are following the rules of your new reality perfectly. To fix you, you don't need to fix your eyes; you need to relearn the traffic laws.

Technical Summary

1. Problem Statement

The paper addresses a critical gap in current computational accounts of delusions (bizarre, fixed beliefs in neuropsychiatric conditions like schizophrenia). While Predictive Coding (PC) frameworks are dominant in modeling belief formation, existing models struggle to explain two core features of delusions:

1. Thematic Recurrence: Why delusions often follow specific, recurring themes.
2. Imperviousness to Evidence: Why individuals maintain these beliefs despite contradictory sensory evidence.

Standard PC models are often "memoryless," resetting beliefs randomly for each stimulus. In these models, increasing the precision of priors (a common hypothesis for psychosis) would theoretically lead to unstable, shifting beliefs rather than fixed delusions. Furthermore, standard models suggest that holding a false belief against evidence should result in high prediction errors (uncertainty), yet delusional patients often report a sense of certainty and reduced anxiety. The authors propose that current models fail because they treat delusions as faulty inference rather than a consequence of a reorganized generative model.

2. Methodology

The authors utilized Hybrid Predictive Coding, a framework combining iterative inference with amortised inference (a fast, feedforward mapping from sensory input to belief states).

Model Architecture

• Structure: A four-layer hierarchical network trained on the EMNIST dataset (handwritten digits 0–5 and letters A, B, C, D, E, S).
• Components:
  1. Inference Network: Performs iterative belief updating via top-down feedback and bottom-up prediction errors.
  2. Two Amortisation Networks: Context-sensitive networks that provide an initial "best guess" (feedforward sweep) for the inference network. One network specializes in digits, the other in letters.
• Experimental Setup:
  1. Progenitor Model: Trained on both digits and letters with correct context mapping (digits $\to$ digit network, letters $\to$ letter network) and supervised top-level labels.
  2. Twin Models: The progenitor weights were cloned into two models:
     • Control Twin: Retained correct amortisation mappings.
     • Delusional Twin: The amortisation networks were swapped (e.g., the digit network processed letter inputs). This simulated an "inappropriate initialisation" of beliefs.
  3. Self-Supervised Phase: Both twins were trained further where the top-level beliefs were clamped to the output of the (now incorrect) amortisation network. This simulated "increased precision" of high-level priors, forcing the model to accept the initial (incorrect) guess as ground truth.

Simulation Phases

1. Training: The progenitor learned to reconstruct inputs and classify them.
2. Disruption: The delusional twin began with incorrect initial beliefs due to the swapped amortisation.
3. Reorganisation: The model was forced to minimize prediction error while maintaining the clamped, incorrect high-level beliefs.
4. Recovery Test: Models were subjected to a "self-sustaining" phase (no clamping) to see if they could revert to non-delusional states after varying durations of the delusional phase.

3. Key Contributions

• Reframing Delusions: The paper argues that delusions are not the result of "faulty inference" (bad math) but are adaptive consequences of a generative model reorganizing itself to accommodate atypical initial conditions (inappropriate amortisation + high precision).
• Generative Model Reorganisation: The study demonstrates that the brain (or model) can structurally alter its internal weights to make a delusional belief the most likely explanation for sensory data, thereby eliminating prediction error.
• Amortisation Trap: The authors introduce the concept of an "amortisation trap," where a failure in the initial feedforward sweep, combined with rigid high-level priors, forces the system to learn a new causal structure that validates the error.
• Mechanism for Fixity: The model explains why delusions are impervious to evidence: once the generative model reorganizes, the "delusional" belief generates the lowest prediction error, making it the optimal inference within that specific model.

4. Key Results

A. Emergence of Delusional Beliefs

• Initial Phase: When the delusional twin first encountered mismatched inputs (e.g., a '4' processed by the letter network), it initially misclassified the input (as 'A'). However, iterative inference could still correct this error, leading to the correct classification ('4').
• Reorganisation Phase: As training continued with clamped high-level beliefs, the model could no longer correct the error. Instead, the generative model reorganized.
  • The model learned to map the visual features of a '4' to the belief state 'A'.
  • Result: The delusional twin consistently misclassified '4' as 'A' even when inference started from random states (not just amortised ones). This proved the generative model itself had changed, not just the initial guess.

B. Prediction Error Dynamics

• Transient Spike: Initially, prediction error spiked because the clamped belief ('A') conflicted with the input ('4').
• Normalization: Once the generative model reorganized, prediction error dropped to control levels. The model successfully explained the input '4' using the belief 'A' without increasing uncertainty.
• Implication: This explains why delusions feel certain to the patient; the internal model is perfectly consistent, even if it is objectively wrong.

C. Attractor Points and Recovery

• New Attractors: The reorganization created a new "attractor point" in belief space where 'A' was the stable state for input '4'.
• Duration Dependence: The ability to recover from delusions was inversely proportional to the duration of the delusional phase.
  • Brief exposure: The model could quickly remap and recover.
  • Extended exposure: The gradients supporting the "healthy" model eroded, and the "delusional" model was reinforced. Recovery became partial or impossible, even when the high-level precision constraint was removed.

D. Representational Alignment

• Analysis using Centered Kernel Alignment (CKA) showed that the delusional twin's representations diverged significantly from the progenitor model, particularly in the higher layers (Layer 3), while lower layers remained relatively faithful to the input. This indicates a top-down restructuring of the world model.

5. Significance and Implications

• Theoretical Shift: The paper challenges the view that delusions are simply "over-weighted priors." Instead, it posits that delusions are rational inferences within a reorganized, idiosyncratic generative model. The "bizarreness" is only apparent to an observer with a different generative model.
• Clinical Implications:
  • Early Intervention: Since the "non-delusional" generative model erodes over time, early intervention is critical to prevent the structural reorganization of the brain's feedback connectivity.
  • Therapeutic Strategy: Treatment should not focus solely on challenging the content of the belief (inference) but on re-training the generative model itself.
  • Relapse Risk: Because the healthy model is rarely completely erased, "delusional" structures may remain latent, explaining why patients are prone to relapse even after periods of remission.
• Cultural Context: The findings suggest that generative models are shaped by external supervision (parents, culture). This offers a computational explanation for why displaced persons or those in conflicting cultural environments are at higher risk for psychosis (adapting to a different shared generative model).

In summary, the paper provides a formal, computational mechanism showing how contextual failures in belief initialization, combined with high precision, drive a structural reorganization of the brain's generative model, resulting in the fixed, certain, and theme-specific nature of delusions.