Clean Canvas Principle
Clean Canvas Principle
Status: active β experimentally confirmed Confidence: high (four experiments, adversarial review: Grok + Thoth GND) Domain: machine learning / computational experiments Source: collab-thoth experiments Exp 3a/3b/3c/4a (2026-10) Note: Interpretive connections (Laplacian, ONM, biological parallels) are in the companion blog post, not here. This article is restricted to experimental findings.
Core Finding
Training a neural network on white Gaussian noise produces significantly better out-of-distribution (OOD) generalisation than training on structured noise from another networkβs processing (βexhaustβ). With the same frozen weights, adding structured exhaust at inference time provides a small but consistent OOD benefit; white noise at inference hurts.
What We Know
Finding 1 β The Clean Canvas Effect (Exp 3a, confirmed)
Seven training canvas conditions were tested across 5 seeds. Primary metric: mean OOD accuracy across 30Β°, 45Β°, and 60Β° rotation on CIFAR-10 (pre-registered).
| Training canvas | Mean OOD | vs WHITE |
|---|---|---|
| WHITE (Gaussian IID) | 30.33% | β |
| SHUFFLED-EXHAUST | 29.56% | β0.77pp |
| PINK (1/f noise) | 28.70% | β1.63pp |
| RANDOM-LABEL-EXHAUST | 28.66% | β1.67pp |
| SAME-EXHAUST | 28.09% | β2.24pp |
| DEAD-EXHAUST | 28.05% | β2.28pp |
| CROSS-EXHAUST | 28.03% | β2.30pp |
- Kill test (WHITE vs CROSS-EXHAUST): Ξ = +2.30pp, t = 5.83, CI [+1.20, +3.39], p = 0.004.
- In-distribution accuracy (~83%) is identical across all arms β the canvas affects OOD only.
- Mechanism is not established by this experiment; the data show an effect, not a cause.
Finding 2 β Inference Canvas Effect (Exp 3b corrected, replicated)
With frozen WHITE-trained weights, eight inference canvases were tested. All numbers are mean OOD across three angles (pre-registered metric).
| Inference canvas | Ξ vs NO-NOISE | t | Seeds |
|---|---|---|---|
| CROSS-EXHAUST | +0.55pp | 10.45 | 5/5 |
| SAME-EXHAUST | +0.52pp | 11.97 | 5/5 |
| RANDOM-LABEL | +0.34pp | 5.48 | 5/5 |
| SHUFFLED | +0.31pp | 4.13 | 5/5 |
| NO-NOISE | 0 | β | β |
| PINK | β0.28pp | β7.35 | 0/5 |
| WHITE | β0.60pp | β24.39 | 0/5 |
Effect is real, consistent, and small (~0.5pp, ~1.7% relative improvement). The inference effect is approximately 4Γ smaller than the training effect. Not a Law β a finding.
PINK at inference (β0.28pp, 0/5 seeds): PINK noise hurts. The mechanism is not established by this experiment. For discussion of a principled hypothesis (Schroederβs system-matched noise concept), see the companion blog post.
Finding 3 β Canvas Dependency (Exp 3c, exploratory)
All combinations of 3 training canvases Γ 8 inference canvases (360 evaluations).
| Matched pair | Ξ vs own NO-NOISE | t |
|---|---|---|
| WHITE-trained + WHITE-infer | β0.61pp | β8.96 |
| CROSS-trained + CROSS-infer | +0.57pp | +8.54 |
| SAME-trained + SAME-infer | +0.63pp | +17.02 |
White-trained models are hurt by white noise at inference. Exhaust-trained models are helped by their training canvas at inference. The mechanism is not established. Flagged exploratory (n=5 insufficient for the full 3Γ8 interaction matrix).
Finding 4 β Dose-Response (Exp 4a, confirmed)
Sweeping inference noise intensity (Ο) from 0 to 0.20 for SAME-EXHAUST on WHITE-trained weights (9 levels Γ 5 seeds Γ 3 rotations = 135 evaluations).
| Ο | Mean OOD | Ξ vs Ο=0.05 |
|---|---|---|
| 0.00 | 30.56% | β0.56pp |
| 0.05 | 31.12% | reference |
| 0.07 | 31.21% | +0.09pp |
| 0.10 | 31.24% | +0.12pp |
| 0.15 | 31.03% | β0.09pp |
| 0.20 | 30.67% | β0.45pp |
Verdict: H0 retained. Ο=0.05 is near-optimal. The highest observed mean (Ο=0.10, +0.12pp) is within noise: 3/5 seeds, t=1.41, well below the pre-registered threshold of Ξβ₯+0.30pp. The dose-response follows an inverted-U shape (confirmed descriptively). At high Ο the performance drops below baseline β consistent with noise dominating signal, not a specific theoretical prediction.
System-Matched Noise
Schroeder (Fractals, Chaos, Power Laws, 1991) shows that pink noise is the psychoacoustic equivalent of white noise for cochlear systems: equal power per octave is perceptually flat because the cochlea processes frequency logarithmically. This is a property of that specific biological architecture.
The experiments show that white Gaussian noise is the best training canvas for this CNN (spatial pixel processor). Whether this is because white noise is βsystem-matchedβ for a CNN in the same sense that pink is system-matched for a cochlea is a hypothesis β not established by these experiments. The Schroeder framing is discussed in the companion blog post.
What the data establish: white Gaussian noise produces the best OOD performance as a training canvas. PINK at inference hurts. The reasons are not identified here.
Open Questions
| ID | Question | Priority |
|---|---|---|
| Q-CC-01 | Can the Clean Canvas Principle be formalised mathematically? Current framing is suggestive, not a derivation. | HIGH |
| Q-CC-02 | Does the inference canvas effect extend to architectures other than SmallResNet or datasets other than CIFAR-10? | HIGH |
| Q-CC-03 | What is the mechanism for the training canvas effect? Mechanism was not measured β only the performance outcome. | HIGH |
| Q-CC-04 | Does the system-matched noise hypothesis (Schroeder) correctly predict the PINK result, and does it generalise? | MED |
| Q-CC-05 | Does the dose-response peak position (Ο*β0.07β0.10) depend on B-model quality or only on Aβs architecture? | MED |
Experimental Archive
All code, data, and reviews: github.com/nagapi2357-ui/clean-canvas-principle
| Experiment | What | Result |
|---|---|---|
| Exp 3a | 7 training canvases, 5 seeds | WHITE wins by 2.30pp (t=5.83, confirmed) |
| Exp 3b corrected | 8 inference canvases, frozen WHITE weights | Structured exhaust helps (+0.52pp SAME), white hurts (β0.60pp) |
| Exp 3c | 3Γ8 canvas-swap matrix | Canvas dependency confirmed directionally (exploratory) |
| Exp 4a | Ο-sweep, 9 levels | Ο=0.05 near-optimal, inverted-U descriptively confirmed, H0 retained |
Adversarial reviews: Grok (xAI, metric labeling corrected) and Thoth GND (effects confirmed from raw files, framing corrected).
Cross-links: [[biological-resonance]] [[neural-resonance]] [[entropy-composite-smoothing]] [[prime-composite-duality]] [[ontological-number-map]] (Interpretive connections to these topics are in the companion blog post, not established by the experiments.)