diff options
| -rw-r--r-- | MIXED_TRAFFIC.md | 9 | ||||
| -rw-r--r-- | RESULTS.md | 12 | ||||
| -rw-r--r-- | REVIEW_SCORECARD.md | 1 | ||||
| -rw-r--r-- | ROADMAP.md | 6 |
4 files changed, 28 insertions, 0 deletions
diff --git a/MIXED_TRAFFIC.md b/MIXED_TRAFFIC.md index 4852ef0..0412bd5 100644 --- a/MIXED_TRAFFIC.md +++ b/MIXED_TRAFFIC.md @@ -129,6 +129,15 @@ and cost invariants therefore pass, but they do not rescue the failed finite, accuracy, alignment, tracking, or norm checks. MT-2 and MT-3 remain untouched; the branch closes without changing traffic strength or predictor schedule. +A post-failure training-only audit performs no validation or test evaluation +and stops at the first training-active nonfinite tensor. The first failures are +the stem forward weight and its momentum at task step 23 for raw, 69 for +norm-matched raw, and 74 for innovation. The predictor slope and reciprocal +feedback weights remain finite at those points. Norm matching and innovation +therefore delay the same multiplicative local instability rather than remove +it. `THEORY.md` derives the linearized residual-coupling term `D W C` and shows +why residual RMS alone cannot certify stability. + ## MT-2: frozen full validation gate Copy MT-1 exactly for 200 epochs. The three jobs remain seed 0 and do not touch @@ -937,6 +937,18 @@ alignment, feedback-tracking, and usable matched-norm checks fail. MT-2 and the five-seed test confirmation remain untouched. This closes the controlled standard-ResNet accept recovery and leaves the reviewer estimate at 5/10. +The clean training-only failure-localization records at revision `e996231` +evaluate neither validation nor test data. They identify the first +training-active nonfinite tensors as stem `W[0]` and its momentum at step 23 +for raw, step 69 for norm-matched raw, and step 74 for innovation. Predictor +slopes, feedback weights, and feedback momenta remain finite when each +diagnostic stops. At innovation step 1, loss is +2.91 and used-signal RMS is 0.00477; by step 16 they reach `4.14e5` and 1.76. +Thus norm matching and subtraction substantially delay, but do not stabilize, +the forward local feedback mode. The linearized update in `THEORY.md` exposes +the multiplicative term `D W C` and proves that any positive effective scalar +mode has momentum transition spectral radius above one. + ## How to run `experiments/run.py --mode {bp,fa,dfa,sdil} --dataset {mnist,fmnist,cifar10} --depth D --residual {0,1} --act {tanh,gelu,silu,relu}` Batteries: `experiments/run_v2.sh <ds> "<depths>" <res> <act> "<seeds>" <ep> <pfx>`. diff --git a/REVIEW_SCORECARD.md b/REVIEW_SCORECARD.md index bbc5bd3..035448f 100644 --- a/REVIEW_SCORECARD.md +++ b/REVIEW_SCORECARD.md @@ -102,6 +102,7 @@ retroactively reopened by success on standard vision benchmarks. | 2026-07-22 / `2ef7f94` modified KP-2 | One clean full ResNet record reaches 91.26%, 0.36 points below matched BP, with 0.9994 early alignment, 0.9997 late feedback cosine, zero queries, and 1.326x BP MACs | 5 → 5 | Removes substrate stability as the immediate blocker and opens the load-bearing innovation experiment, but prior-art reciprocal plasticity earns no novelty credit | | 2026-07-22 / mixed-traffic MT-0 | Exact raw/matched/innovation mechanics pass; synthetic ResNet-20 matched batch peaks at 0.857 GB allocated on GTX 1080 without task data | 5 → 5 | Removes implementation, graph, and memory objections before endpoint access; supplies no evidence yet that innovation is useful | | 2026-07-22 / `f310ad5` mixed-traffic MT-1 | All three frozen conditions become nonfinite in epoch 1 and end at 10%; ratio calibration, predictor warmup, zero-query, and cost checks pass | 5 → 5 | Fails to make innovation load-bearing on a standard ResNet and closes MT-2/MT-3; the prior mechanism-only evidence survives but empirical support does not improve | +| 2026-07-22 / `6531aed`, `d818fc2` MT-1 diagnosis | Training-only localization finds the first active failure at stem steps 23/69/74 for raw/matched/innovation; residual coupling yields the verified multiplicative operator `D W C` | 5 → 5 | Sharpens the failure into an operator-stability requirement and improves soundness of the negative analysis, but adds no successful held-out SDIL endpoint | Future rows are appended only after an audited frozen stage. A score staying flat is informative: engineering, theory exposition, or visualization may make the paper more defensible without @@ -502,6 +502,12 @@ checks pass while all performance, finite, alignment, and tracking checks fail. Per the frozen stop rule, MT-2 and MT-3 remain untouched, no weaker traffic ratio is allowed, and this standard-ResNet accept path closes at 5/10. +The training-only post-failure audit localizes the common instability to stem +`W[0]` and its momentum at steps 23/69/74 for raw/matched/innovation, +respectively, with no validation or test evaluation. This narrows any future +independent method to operator-level gain control or dissipativity; more +endpoint tuning of the same residual-RMS gate is not justified. + Prepare convolutional local-update primitives and ResNet-20/32/56 protocols early. Queue frozen runs opportunistically on authorized idle GPUs. Because BurstCCN already reports CIFAR-10 and ImageNet scaling, dataset scale alone is not novel. The oral-level target is a memorable joint |