# Эксперимент: Distinct-kink complexity regularizer and merger (#1387)

{
  "worked": true,
  "confidence": 8,
  "verdict": "Implemented canonical ReLU kink normalization and exact same-hyperplane coefficient merging. The algebra check and dense-probe comparison showed only float32-level discrepancies (maximum about 1.7e-6 and 1.4e-6 respectively), while merging reduced width from 48 to 24 and parameters from 147 to 75 without changing validation MSE or mean Jacobian norm. This verifies the claimed exact merger phenomenon, but does not demonstrate that a differentiable regularizer discovers duplicates spontaneously during training.",
  "metrics": {
    "baseline": "Redundant width-48 coefficient-split model: validation MSE 0.0195736885, mean Jacobian norm 1.35254955, 147 parameters.",
    "idea": "Merged width-24 model: validation MSE 0.0195736848, mean Jacobian norm 1.35254955, 75 parameters; 2.0x width reduction and probe RMSE function error 2.11e-7."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 kink_experiment.py",
  "files": [
    "kink_experiment.py",
    "results/metrics.json"
  ],
  "limitations": "The experiment uses deliberately constructed duplicate neurons, so it isolates exact merger correctness rather than testing duplicate formation under regularization. It does not test CIFAR-10, magnitude/L0 pruning baselines, approximate clustering tolerance, inference latency, or robustness under adversarial perturbations."
}