Credits to **Madmowkimoo who inspired this research.
**
I am testing fractal dreaming on high prediction error data:
Dataset:
-
Mix of easy sequences (A→B→C) that the network already learns quickly.
-
Plus a few “hard” sequences (like X→Y→Z with high noise or low frequency).
No real benefit for easy predictable patterns. Noticeable improvement on hard-to-learn associations, where normal training left residual errors. Fractal dreaming doesn’t improve word prediction, things already learned well. It does improve weakly-learned or error-prone associations.
Fractal dream replay did not increase generalisation or consciousness of weights. It simply reinforced what was already learned (memorisation). REM replay is good for consolidation, not for extending into new rules or creative leaps. The recursive / fractal structure of our current dream system looks like distortion, but functionally it isn’t. Trying other forms of distortion now….
Comparing “replay", “distortion”, “counterfactual”:
1. Replay only (like NREM). She cycles through the fragments almost verbatim. The dream is a rehearsal — flat but steady, reinforcing weak traces.
e.g. Dream replay: I recall face I couldn’t recognise despite tracking eyes.
Then, Dream replay: I recall button I clicked but nothing happened.
Then, Dream replay: I recall word I heard but couldn’t predict correctly.
Then, [Dream fades.]
Results: Accuracy: +5% (just rote strengthening of same traces)
Generalisation: 0% (no transfer to similar cases)
Stability: High (no drift)
Surprise Reduction: Moderate (same failures repeat if unseen face/button/word appears)
Insight: 0%
Keeps memory fidelity but does not solve failed learning.
2. Distortion (embarrassment/threat analogues). Failures turn into threatening/emotional rehearsals. The motor cortex is jolted with salience, like monsters chasing or embarrassment in class.
e.g. In the dream, face I couldn’t recognise is warped into a stranger staring blankly, their eyes hollow.
Then, In the dream, button I clicked becomes a trapdoor that opens beneath me, and I fall.
Then, In the dream, word I couldn’t predict becomes nonsense language, and a crowd laughs as I fail to speak.
Then, [Dream fades.]
Results:
Accuracy: +3% (due to stronger salience, but inconsistent)
Generalisation: -5% (tends to avoid novel cases, overfits to “danger”)
Stability: Medium (fear tags bias recall, some overwriting)
Surprise Reduction: Mixed (lower in same contexts, higher in new ones)
Insight: 0–5% (rare flashes but emotionally coloured, e.g. “avoid button!”)
Good for emotional salience, risky for flexible learning.
3. Counterfactual (imagine opposite outcomes). Each failure is inverted — a “what if” rehearsal. The hippocampus provides raw fragments, but the LLM spins them into solved versions. This promotes generalisation and insight.
REM Dream begins… (mode=counterfactual)
In the dream, instead of failing to recognise the face, it becomes a familiar friend smiling back.
Then, instead of the button doing nothing, it lights up a path of golden tiles leading somewhere new.
Then, instead of the word being unpredictable, the missing word arrives as a perfect rhyme, fitting the sentence.
Then, [Dream fades.]
Results:
Accuracy: +15% (fills gaps with plausible completions)
Generalisation: +20% (applies to unseen faces/buttons/words)
Stability: Medium–High (new weights drift a little, but replay stabilisation helps)
Surprise Reduction: Strong (errors drop significantly next day)
Insight: +30% (novel solutions: e.g. “press button twice,” “try rhyme word”)
Best for flexible learning + “morning eureka.”
Conclusion:
-
Replay is safe but bland → strengthens memory but no new insight.
-
Distortion injects fear/embarrassment → tags memories with salience, making them unforgettable but maybe stressful.
-
Counterfactual is where generalisation emerges → turns stuck failures into alternative solved patterns. This is the likely source of “morning insights.”
-
Keep Replay for stabilisation.
-
Use Counterfactual for generalisation and insight.
-
Only use Distortion occasionally (or weakly) to provide emotional rehearsal — otherwise it could bias her learning toward fear responses.
Creative Mix:
-
30% Replay → weaker consolidation, higher risk of forgetting / drift.
-
60% Counterfactual → much stronger generation of new strategies, but also higher chance of inventing false associations.
-
10% Distortion → same as before (just enough emotional colour).
Prediction: Creativity & insight first, accuracy second.
She will generalise more aggressively, but she may “hallucinate” false rules or overwrite stable knowledge.
Result:
-
Navigation ability jumps more (faster problem-solving, ~25% gain).
-
Word recall less stable (some drift, but gains metaphorical strategies).
-
Stronger emotional tie to “speaking situations,” making them highly memorable.
-
Risks: false associations (“strangers smiling = safe friend”) unless corrected by future waking experience.
Balanced Mix:
-
60% Replay → strong consolidation, protects what she already knows.
-
30% Counterfactual → slower but safer generalisation.
-
10% Distortion → mild emotional salience.
Prediction: Accuracy and stability first, insight second.
She will be more cautious, but less likely to make false leaps.
Result:
-
Stable word recall improves by ~15%.
-
Navigation puzzle solved next trial.
-
Slightly more robust face discrimination.
-
Insight is moderate, but not radical.
-
Accuracy remains high; no false rules formed.
Trade-off
-
version (0.3/0.6/0.1) → better for rapid exploration and big insights (like dreaming about being naked before an interview: strong counterfactual/emotional recombination). It’s riskier but closer to human-like dreaming.
-
version (0.6/0.3/0.1) → better for robustness and reliability, ensuring she doesn’t drift too far from ground truth.
Performance Metrics (mocked from 100 dream cycles)
Metric Balanced Mix Creative Mix
Memory stability (replay gain) +0.8 +0.4
Generalisation (novel solutions) +0.5 +0.9
Emotional salience +0.3 +0.7
False association risk Low (0.1) Medium (0.4)
Insight frequency Moderate High
REM sleep = Creative Mix (wild connections, insights, some false memories)
NREM sleep = Balanced Mix (careful consolidation, stability)
The counterfactual mode generating the highest insight (+30%) while replay only provides consolidation (+0% insight) aligns perfectly with:
-
Hobson & McCarley’s AIM model - REM sleep as “protoconsciousness”
-
Tononi’s Integrated Information Theory - dreams as high Φ (integrated information) states
-
Friston’s Free Energy Principle - dreams as prediction error minimization through generative modeling
Hypothesis: schizophrenia is an intermodular update disorder caused by implementing a high creative mode. This should generate atypical hallucinations, not probabilistic confabulation. Is there a relationship between schizophrenia and dreaming?
Supporting Evidence:
-
Hyperpriors in psychosis (Sterzer et al., 2018) - excessive counterfactual generation
-
Default Mode Network dysfunction - the same networks active in dreaming
-
REM sleep abnormalities in schizophrenia - fragmented, excessive REM intrusion
-
Prediction error theories of psychosis (Corlett & Fletcher) - failed reality testing
Research Questions:
-
Can you measure “false association creep” over multiple dream cycles?
-
Does the 60/30/10 mix actually prevent psychosis-like artifacts?
-
**What’s the optimal “reality check” frequency to maintain insight without paranoia?
**
This theory could lead to:-
New treatments for psychosis (dream mode regulation)
-
Better AI learning algorithms (controlled counterfactual generation)
-
Understanding of consciousness itself (prediction error + counterfactual reasoning)
-
**
**