From 2a6f72e22a814fb4093174a2fb1ca2af56c7e63c Mon Sep 17 00:00:00 2001 From: YurenHao0426 Date: Thu, 23 Jul 2026 07:09:14 -0500 Subject: figures: add audited ResNet confirmation panel --- results/figs/figure4_resnet_confirmation_caption.md | 20 ++++++++++++++++++++ 1 file changed, 20 insertions(+) create mode 100644 results/figs/figure4_resnet_confirmation_caption.md (limited to 'results/figs/figure4_resnet_confirmation_caption.md') diff --git a/results/figs/figure4_resnet_confirmation_caption.md b/results/figs/figure4_resnet_confirmation_caption.md new file mode 100644 index 0000000..2de4e45 --- /dev/null +++ b/results/figs/figure4_resnet_confirmation_caption.md @@ -0,0 +1,20 @@ +# Standard-ResNet confirmation figure caption + +**Figure 4 | Somato-dendritic innovation survives a frozen standard-ResNet +confirmation.** All task endpoints are CIFAR-10 test results from the five +untouched seeds 10--14 on a standard ResNet-20; lines or bands show paired +seeds or mean ± sample standard deviation. **a,** Dynamic neutral-projection +innovation under four-times-RMS soma-predictable apical traffic matches clean +reciprocal Kolen--Pollack credit. Diamonds show means and 95% normal intervals. +Dynamic innovation gains 0.196 accuracy points on average; the predeclared +one-sided 95% upper bound on its deficit is 0.131 points. **b,** Final +layerwise cosine between the raw apical or innovation signal and the exact +negative hidden-state gradient. +Exact gradients are audit-only and never used for learning. Residualization +removes the directionally contaminating traffic while retaining an early-layer +cosine of 0.999687. **c,** Mean reciprocal-feedback tracking across all 200 +epochs. Resource annotations report measured GTX 1080 wall time and peak +allocation, hardware-independent MAC estimates relative to the matched +50,000-example BP accounting reference, and logical task-loss queries. The +paired instruction-off neutral observation is explicitly counted and is not +claimed to be free or single-phase. -- cgit v1.2.3