diff options
Diffstat (limited to 'results/contrastive_bias/B1_SUMMARY.md')
| -rw-r--r-- | results/contrastive_bias/B1_SUMMARY.md | 28 |
1 files changed, 28 insertions, 0 deletions
diff --git a/results/contrastive_bias/B1_SUMMARY.md b/results/contrastive_bias/B1_SUMMARY.md new file mode 100644 index 0000000..19ca57c --- /dev/null +++ b/results/contrastive_bias/B1_SUMMARY.md @@ -0,0 +1,28 @@ +# Dual Prop state-bias screen + +The central result is positive, but the complete pre-registered gate fails. + +In the author Dual Prop miniCNN, neuron-specific activity-dependent bias makes +the raw local update nonfinite in epoch 1 at every tested bias ratio. The +innovation rule stays finite for all 20 epochs and remains close to both clean +Dual Prop and an oracle that subtracts the generated bias exactly. + +| Bias / clean teaching RMS | Raw | Innovation | Oracle | +|---:|---:|---:|---:| +| 0.25 | 9.40, nonfinite | 71.10 | 70.58 | +| 1 | 9.40, nonfinite | 70.18 | 69.80 | +| 4 | 9.40, nonfinite | 68.96 | 69.86 | + +Values are final validation accuracy in percent. Clean Dual Prop is 69.66%. +The largest innovation-to-clean gap is 1.44 points and the largest +innovation-to-oracle gap is 0.90 points. The maximum bias remaining after +subtraction is `8.96e-8` of the raw bias, and the predictor sees zero task +instruction observations. + +The complete B1 gate is `fail` for two reasons unrelated to the recovery +comparison: clean ends 0.34 points below the frozen 70% threshold, and the +common-bias endpoint differs from clean by 2.28 points even though its teaching +difference error is exactly zero. The latter shows that a 0.2-point single-run +accuracy tolerance is not a reliable identity check on these GPU kernels. +Because the complete gate fails, B2 is not opened automatically. A new +confirmation protocol must be frozen before running more seeds. |