Эксперимент: Projected Bures Covariance Pooling (#1344)

{ "worked": false, "confidence": 9, "verdict": "Implemented exact unit-step BW barycenter updates with post-update eigenvalue clipping, plus BW distance, objective, conditioning, and pooling checks. The feasible toy case converged, but projection was inactive because no eigenvalues left the interval; the non-expansiveness check was broadly supportive, with mean distance ratio 0.883 and maximum 1.0000000000000004. The synthetic classification task was saturated at 1.0 accuracy for every method, so no empirical task improvement or meaningful projection benefit was demonstrated.", "metrics": { "baseline": "Arithmetic pooling: accuracy 1.000, mean condition number 9.868, mean BW objective 0.011245; unprojected BW: accuracy 1.000, condition 10.062, objective 0.011204.", "idea": "Projected BW: accuracy 1.000, mean condition number 10.062, mean BW objective 0.011204; projection was inactive in the tested feasible run. Projection distance ratio mean 0.883, maximum 1.0000000000000004." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json" ], "limitations": "No CNN/CIFAR experiment, differentiable PyTorch implementation, GPU benchmark, FLOP/speed measurement, or adversarial transient-floor-exit instance was completed. The random exit search was shortened after exceeding the runtime budget, and the classification setup was too easy to distinguish methods." }