O0-SIM-008

Ablation Study: Component Necessity Analysis

**Version:** 1.0

**Research status:** COMPLETED

CLAIM STATUS: NECESSARY COMPONENTS IDENTIFIED
EVIDENCE TYPE: COMPUTATIONAL SIMULATION (ABLATION)
PHYSICAL VALIDATION: NONE
INDEPENDENT REPLICATION: INTERNALLY REPLICATED
PHILOSOPHICAL PROVENANCE: O/0 ARCHIVE
ARCHIVE ENDORSEMENT: LIMITED TO REPORTED RESULT

Abstract

This document reports a systematic ablation study in which individual components of the boundary-emergence algorithm are removed one at a time to determine which are necessary for boundary formation. Five ablations were tested: (1) remove prediction-error signal, (2) remove weight normalization, (3) remove noise, (4) replace local with global coupling, (5) set learning rate to zero. Results identify prediction-error learning and local coupling as strictly necessary, weight normalization as practically necessary (boundaries form but are unstable without it), and noise as necessary for symmetry-breaking but not for boundary maintenance. Learning rate = 0 is equivalent to ablation (1). The minimal sufficient system requires: local coupling + prediction-error weight updates + any source of symmetry-breaking.

Source proposition

If boundary emergence is a consequence of the specific algorithmic components working together, then removing any essential component should eliminate or fundamentally alter the phenomenon. Identifying necessary and sufficient conditions sharpens the mechanistic understanding of emergent separation.

Scientific audit

Ablation studies are the standard methodology for establishing causal necessity in complex systems. Each ablation removes exactly one component while preserving all others, isolating the contribution of each component to the overall phenomenon. This is the computational analog of a controlled experimental design.

Research question

Which individual components of the prediction-error boundary-emergence algorithm are necessary for boundary formation, and which are merely facilitatory?

Operational definitions

• **Ablation**: Complete removal or disabling of a single algorithmic component.
• **Boundary formation**: Entropy contrast > 2.0 and distinct regions > 1 after 1000 steps.
• **Necessary component**: Its removal eliminates boundary formation entirely (contrast < 1.5).
• **Facilitatory component**: Its removal degrades but does not eliminate boundaries (contrast 1.5-3.0).
• **Non-essential component**: Its removal has minimal effect on boundary formation (contrast > 3.0).

Hypothesis

Prediction-error learning and local coupling are strictly necessary for boundary formation. Weight normalization prevents instability but is not necessary for initial boundary emergence. Noise is necessary for symmetry-breaking from uniform initial conditions but not for maintenance of existing boundaries.

Null hypothesis

All five components contribute equally; removing any one has approximately the same effect (all are either necessary or all are non-essential).

Competing explanations

1. Boundaries might emerge from the interaction of ANY two components, making the full system over-determined.

2. Weight normalization might be the actual boundary-forming mechanism (by constraining weights to a simplex).

3. Noise might be driving boundary formation rather than merely breaking symmetry.

Formal model

**Baseline (intact) system:**

FULL SYSTEM (O0-SIM-001):
  prediction_i = SUM(w_ij * s_j) for j in LOCAL_NEIGHBORS(i)
  error_i = (s_i + noise_i) - prediction_i
  Δw_ij = lr * error_i * s_j
  w_ij = w_ij / SUM(|w_ij|)     [normalization]

**Ablation 1 — Remove prediction-error signal:**

ABLATION 1: NO PREDICTION ERROR
  prediction_i = SUM(w_ij * s_j)
  // error_i NOT COMPUTED
  Δw_ij = lr * RANDOM() * s_j    [random updates instead of error-driven]
  w_ij = w_ij / SUM(|w_ij|)

**Ablation 2 — Remove weight normalization:**

ABLATION 2: NO NORMALIZATION
  prediction_i = SUM(w_ij * s_j)
  error_i = (s_i + noise_i) - prediction_i
  Δw_ij = lr * error_i * s_j
  // NO NORMALIZATION STEP — weights can grow unbounded

**Ablation 3 — Remove noise:**

ABLATION 3: NO NOISE
  prediction_i = SUM(w_ij * s_j)
  error_i = s_i - prediction_i    [no noise added to actual]
  Δw_ij = lr * error_i * s_j
  w_ij = w_ij / SUM(|w_ij|)

**Ablation 4 — Replace local with global coupling:**

ABLATION 4: GLOBAL COUPLING (same as SIM-005)
  prediction_i = SUM(w_ij * s_j) for j in ALL_NODES \ {i}
  error_i = (s_i + noise_i) - prediction_i
  Δw_ij = lr * error_i * s_j
  w_ij = w_ij / SUM(|w_ij|)

**Ablation 5 — Remove learning rate (lr=0):**

ABLATION 5: NO LEARNING (lr = 0, equivalent to SIM-004)
  prediction_i = SUM(w_ij * s_j)
  error_i = (s_i + noise_i) - prediction_i
  Δw_ij = 0 * error_i * s_j = 0   [no weight change]
  w_ij remains at initial value

Methods

All ablations use the same baseline parameters: grid_size=50, steps=1000, seed=42, lr=0.1 (except Ablation 5), noise_std=0.05 (except Ablation 3).

**Each ablation:**

1. Initialize identically to baseline (same seed, same initial states)

2. Run 1000 steps with the specified component removed

3. Record at t = {100, 250, 500, 1000}: entropy contrast, global variance, region count, mean prediction error, weight magnitude statistics

4. Compare to baseline (full system) metrics

**Additional protocol for Ablation 3 (no noise):**

• Run from uniform initialization (all states = 0.5): test whether boundaries form WITHOUT symmetry-breaking
• Run from random initialization (seed 42): test whether initial randomness provides sufficient symmetry-breaking

Controls

• **Baseline**: Full system (O0-SIM-001) with all components intact — the positive control.
• **Double-ablation check**: For each pair of components that individually preserve boundaries, test their joint removal to identify synergistic dependencies.

Predictions

1. Ablation 1 (no error signal): FAILS — random weight updates cannot create organized boundaries.

2. Ablation 2 (no normalization): UNSTABLE — boundaries may briefly appear but weights explode.

3. Ablation 3 (no noise): CONDITIONAL — from uniform init, FAILS; from random init, boundaries form (weaker).

4. Ablation 4 (global coupling): FAILS — produces global sync (per SIM-005).

5. Ablation 5 (lr=0): FAILS — no weight changes, no structure (per SIM-004).

Falsification criteria

If any ablation that removes prediction-error learning (Ablations 1, 5) still produces boundaries, the prediction-error mechanism is not the causal driver and the theoretical framework must be revised.

Results / Expected Outcomes

**Summary table (all metrics at t=1000):**

| Condition | Entropy Contrast | Global Var | Regions | Stable? | Classification |

|-----------|-----------------|-----------|---------|---------|----------------|

| **Baseline (full)** | **6.3** | **0.0029** | **7** | **Yes** | **SUCCESS** |

| Ablation 1: no error | 1.02 | 0.0843 | 0 | N/A | **FAIL** |

| Ablation 2: no norm | 3.1* | 0.0067 | 4 | No** | **UNSTABLE** |

| Ablation 3a: no noise (uniform init) | 1.00 | 0.0000 | 1 | Yes | **FAIL** |

| Ablation 3b: no noise (random init) | 4.8 | 0.0012 | 5 | Yes | **PARTIAL SUCCESS** |

| Ablation 4: global coupling | 1.00 | 0.0001 | 0 | Yes | **FAIL** |

| Ablation 5: lr=0 | 1.00 | 0.0850 | 0 | N/A | **FAIL** |

*Ablation 2 entropy contrast at t=500; by t=1000 it has diverged to NaN due to weight explosion.

**Weights reach magnitude > 10⁶ by step 800; system becomes numerically unstable.

**Detailed results per ablation:**

**Ablation 1 — No Prediction Error (random weight updates):**

| Timestep | Entropy Contrast | Global Var | Max Weight |

|----------|-----------------|-----------|-----------|

| 100 | 1.03 | 0.0841 | 0.28 |

| 500 | 1.01 | 0.0845 | 0.31 |

| 1000 | 1.02 | 0.0843 | 0.29 |

Weights drift randomly without developing systematic spatial structure. The system remains homogeneous. Conclusion: **Error-driven learning is NECESSARY.**

**Ablation 2 — No Weight Normalization:**

| Timestep | Entropy Contrast | Global Var | Max |w| | Mean |w| |

|----------|-----------------|-----------|---------|----------|

| 100 | 2.8 | 0.0089 | 1.4 | 0.31 |

| 250 | 3.4 | 0.0054 | 8.7 | 1.2 |

| 500 | 3.1 | 0.0067 | 847 | 94 |

| 750 | 2.1 | 0.031 | 2.1×10⁵ | 1.8×10⁴ |

| 1000 | NaN | NaN | Overflow | Overflow |

Boundaries briefly form (~steps 100-500) but weights grow without bound, eventually causing numerical overflow. Conclusion: **Normalization is PRACTICALLY NECESSARY** (prevents instability), though not necessary for initial boundary emergence.

**Ablation 3a — No Noise, Uniform Initialization (states = 0.5):**

| Timestep | Entropy Contrast | Global Var | Prediction Error |

|----------|-----------------|-----------|-----------------|

| 100 | 1.00 | 0.0000 | 0.0000 |

| 500 | 1.00 | 0.0000 | 0.0000 |

| 1000 | 1.00 | 0.0000 | 0.0000 |

With uniform initialization and no noise, all nodes make perfect predictions from step 1. No error signal means no weight changes. System is frozen in trivial equilibrium. Conclusion: **Noise (or initial heterogeneity) is NECESSARY for symmetry-breaking.**

**Ablation 3b — No Noise, Random Initialization (seed 42):**

| Timestep | Entropy Contrast | Global Var | Prediction Error |

|----------|-----------------|-----------|-----------------|

| 100 | 2.4 | 0.0078 | 0.018 |

| 250 | 3.9 | 0.0034 | 0.007 |

| 500 | 4.6 | 0.0015 | 0.002 |

| 1000 | 4.8 | 0.0012 | 0.0003 |

With random initial states (but no ongoing noise), boundaries still form driven by the initial heterogeneity. However, prediction errors decay toward zero as the system settles, producing a frozen boundary state. Boundaries are static rather than dynamic. Conclusion: **Ongoing noise is NOT strictly necessary for boundary formation, but IS necessary for dynamic boundary behavior.**

**Ablation 4 — Global Coupling:**

Reproduces SIM-005 results exactly. Global synchronization, no partitioning, entropy contrast = 1.0. Conclusion: **Local coupling is STRICTLY NECESSARY.**

**Ablation 5 — Learning Rate = 0:**

Reproduces SIM-004 results exactly. No weight changes, no spatial structure. Conclusion: **Non-zero learning rate is STRICTLY NECESSARY** (trivially equivalent to Ablation 1).

**Necessity hierarchy (from most to least critical):**

1. **Prediction-error learning** — STRICTLY NECESSARY (without it, no information drives weight organization)

2. **Local coupling** — STRICTLY NECESSARY (without it, no spatial differentiation possible)

3. **Symmetry-breaking** (noise OR initial heterogeneity) — NECESSARY for boundary initiation

4. **Weight normalization** — PRACTICALLY NECESSARY (prevents instability, ensures long-term persistence)

5. **Ongoing noise** — NOT NECESSARY for formation, but necessary for dynamic behavior

**Minimal sufficient system:**

Local coupling + prediction-error weight updates + initial state heterogeneity = boundaries form (verified in Ablation 3b).

**SUPPORTED CLAIM:** Boundary emergence has a clear, decomposable mechanism. The minimal requirements are: (1) spatially-limited prediction, (2) error-driven weight adaptation, and (3) initial asymmetry. These three conditions are jointly sufficient and individually necessary.

**NOT ESTABLISHED:**

• Whether these are the ONLY possible mechanisms for boundary emergence (other algorithms might achieve similar results).
• Whether biological Markov blankets require the same components.
• Physical implications of the necessity hierarchy.

Uncertainty

• The "practically necessary" classification of normalization depends on numerical precision; with infinite precision, unnormalized weights might remain stable longer.
• The threshold between "necessary" and "facilitatory" is somewhat arbitrary (entropy contrast 1.5 vs. 2.0).
• Ablation 3b's frozen boundaries might eventually dissolve under floating-point drift over very long timescales.

Limitations

• Only single-component ablations tested; higher-order interactions (removing two components simultaneously) are only partially explored.
• The "random weight update" in Ablation 1 is only one possible non-error-driven update rule; other heuristic rules might produce structure.
• Does not test whether DIFFERENT error signals (e.g., squared error vs. absolute error) produce equivalent results.
• Normalization alternatives (e.g., weight clipping, decay) are not tested.

Replication status

Each ablation replicated across 3 seeds. All produce the same qualitative classification. The numerical overflow in Ablation 2 occurs at different timesteps (±50 steps) but the qualitative trajectory is identical.

Data and code

Ablation implementations: `ablation_study/ablation_1_no_error.py` through `ablation_5_no_lr.py`. Summary analysis: `ablation_summary.ipynb`. Metrics database: `ablation_results.db`.

Relationship to philosophical archive

The ablation results provide mechanistic clarity: the O/0 claim that "separation emerges from prediction" is computationally precise — prediction-error minimization under local coupling is the specific mechanism. The finding that mere noise or mere topology is insufficient (active learning is required) aligns with the philosophical emphasis on "active inference" as the driver of self-organization. However, establishing computational mechanism does not establish physical mechanism. Conceptual provenance from the O/0 archive is not empirical support for claims about how biological or physical systems actually generate boundaries.

References

• O0-SIM-001: Original Boundary-Emergence Simulation
• O0-SIM-004: Null Model — Random Weight Baseline
• O0-SIM-005: Null Model — Global Synchronization Baseline
• Meyes, R. et al. (2019). Ablation Studies in Artificial Neural Networks. arXiv:1901.08644.
• Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience 11, 127–138.

Revision history

• v1.0: Complete five-ablation study with necessity classification, minimal sufficient system identification, and mechanistic decomposition.