Эксперимент: Extreme-Subset Adversarial Dropout (#1312)

{ "worked": false, "confidence": 9, "verdict": "Built experiment.py with exhaustive least-singular-value subset scoring, a Gaussian theorem scaling check, and a matched toy downstream regression comparison. The extreme-conditioning phenomenon was clearly observed: at N=12,m=6 the minimum score was 0.00021 versus median 0.225, while gamma=2 log(M)/m values were broadly consistent with the theoretical limit -1.386. However, adversarial training worsened identical-pool evaluation, so this MVP does not demonstrate the promised robustness benefit.", "metrics": { "baseline": "Random 20-subset training: final MSE 1.289; exhaustive-mask mean MSE 1.311; worst-1% MSE 2.131; worst MSE 2.294.", "idea": "Lowest-singular-score 20-subset training: final MSE 1.076; exhaustive-mask mean MSE 1.733; worst-1% MSE 2.658; worst MSE 2.852. Extreme score 0.000211 versus median 0.2249, ratio 0.00094." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json" ], "limitations": "Small synthetic regression with N=12,m=6 and fixed candidate subsets, not CIFAR-10 or a deep network. Scores used a fixed random feature matrix rather than a refreshed Jacobian; annealing, swap search, multiple seeds, Bernoulli dropout, and FLOP/accuracy studies were not tested." }