diff options
Diffstat (limited to 'ICLR_2027_REPLAN.md')
| -rw-r--r-- | ICLR_2027_REPLAN.md | 32 |
1 files changed, 22 insertions, 10 deletions
diff --git a/ICLR_2027_REPLAN.md b/ICLR_2027_REPLAN.md index d5444b6..dbd74f4 100644 --- a/ICLR_2027_REPLAN.md +++ b/ICLR_2027_REPLAN.md @@ -29,6 +29,12 @@ locally identifiable operation beats strong bias-specific corrections on a measured physical model and transfers unchanged across local-learning backbones. +SDIL is positioned as a local debiasing plug-in for existing two-state +learners, not as a new optimizer competing on an accuracy--cost frontier. All +comparisons therefore use a matched training/observation budget where +possible, and disclose unmatched phase, storage and arithmetic overhead in a +table rather than a Pareto figure. + ## What the paper must and must not claim The paper may claim that structured differential bias does not disappear with @@ -89,8 +95,9 @@ Pass conditions: - SDIL closes a substantial fraction of the raw-to-oracle gap on every identifiable state-dependent-bias setting; - SDIL beats constant calibration when the bias varies with local state; -- SDIL is nondominated with overclamping on error/cycle span versus charged - observation and circuit cost; +- under a predeclared matched observation/phase budget, SDIL is competitive + with overclamping on error floor and cycle span; unmatched implementation + overhead is reported separately; - common-mode bias gives no artificial SDIL advantage; - task leakage and neutral-to-task shift fail in the direction predicted by the identifiability theory. @@ -137,8 +144,8 @@ Pass conditions: - neutral predictability, rather than corruption RMS, predicts recovery; - SDIL improves every biased raw endpoint and beats constant calibration on state-dependent bias; -- SDIL is competitive with each family's strongest correction after cost is - charged; +- SDIL is competitive with each family's strongest correction under the + matched protocol, with any observation/storage overhead disclosed; - the frozen adapter passes the executable BP-free audit in every family; - clean/common-mode controls show that the gain is not ordinary regularization or a changed optimizer. @@ -176,14 +183,18 @@ appendix or is removed. 1. **Problem and real evidence:** Harnett residual motivation, the two-state measurement model, and the released physical error plateau/cycle drift. 2. **Mechanism on the physical model:** matched noise versus bias; raw, - calibration, overclamping, SDIL and oracle; error/cycle span against charged - cost. + calibration, overclamping, SDIL and oracle; error floor, drift and cycle + span under the matched protocol. 3. **Scaling across backbones:** performance gap and residual bias versus depth, width, biased-block count and task difficulty for physical learning, DP and EP/CpL. 4. **What makes recovery possible:** neutral predictability, drift rate, - task leakage and distribution shift, followed by an accuracy--cost Pareto - summary. + task leakage and distribution shift, showing the predicted success and + failure boundary of the plug-in. + +Phase, observation, storage, arithmetic, wall-time and memory overhead appear +in one compact audit table. They are controls against an unfair comparison, +not a separate Pareto claim. The key scaling plot must show both the failure term and its removal. Plotting only final accuracy across larger clean architectures is not evidence for the @@ -209,7 +220,8 @@ operation, real measured evidence, causal controls and cross-backbone transfer. - extend EP and coupled-learning DCHNs across FashionMNIST, SVHN and CIFAR-10, multiple depths/biased-block counts and five seeds; -- finish five-seed VGG16 DP and the full cost Pareto comparison; +- finish five-seed VGG16 DP and the complete matched-budget bias-recovery + comparison; - cross fixed, state-dependent and slowly drifting bias with matched noise, predictability and neutral/task shift; - keep one frozen predictor configuration across comparable backbones; @@ -269,4 +281,4 @@ Update one evidence table after every completed gate: Every progress report should state positive results, negative results, compute and a fresh reviewer score. Accuracy without the matched strongest baseline, -cost coordinate or frozen protocol does not move the score. +matched protocol or overhead disclosure does not move the score. |