Эксперимент: Jointly Contractive Input-Conditioned RNN (#1241)

{ "worked": false, "confidence": 8, "verdict": "Built an input-conditioned recurrent layer with a learned positive-definite metric, generalized-eigenvalue Lyapunov penalty, pointwise spectral-normalization control, and an explicit product-contraction sanity check. The core math check passed: all basis matrices satisfied the bound and sampled convex-mixture products contracted substantially in the P-norm. However, on the toy 30-step memory task, the Lyapunov model did not improve over controls: long-sequence accuracy was 0.445 for all variants and final MSE was essentially 1.0, so the promised task-level win was not observed.", "metrics": { "baseline": "Free: final MSE 0.99966, long accuracy 0.4453, max basis P-gain 1.2454, runtime 18.32 s; spectral: final MSE 0.99984, long accuracy 0.4453, max basis P-gain 0.9338, runtime 21.05 s", "idea": "Lyapunov penalty: final MSE 0.99953, long accuracy 0.4453, max basis P-gain 0.8815, runtime 21.11 s; math check max basis violation -0.3238 and max normalized sampled product ratio 0.3901 at rho=0.82" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py" ], "limitations": "Only a small synthetic one-bit memory task was tested, with an 8-unit single-layer recurrent model rather than the proposed larger GRU-like architecture or sequential MNIST/copy-add benchmarks. The task setup and rho=0.90 constraint may make 30-step memory intrinsically difficult; no hyperparameter sweep, formal JSR computation, or repeated-seed statistical comparison was performed." }