From a5a04b29bc93bdee4dbce6fa46bfda7cfe843928 Mon Sep 17 00:00:00 2001 From: YurenHao0426 Date: Mon, 27 Jul 2026 14:51:17 -0500 Subject: figure: audit standard-depth SDIL scaling --- results/figs/figure6_standard_depth_scaling_caption.md | 18 ++++++++++++++++++ 1 file changed, 18 insertions(+) create mode 100644 results/figs/figure6_standard_depth_scaling_caption.md (limited to 'results/figs/figure6_standard_depth_scaling_caption.md') diff --git a/results/figs/figure6_standard_depth_scaling_caption.md b/results/figs/figure6_standard_depth_scaling_caption.md new file mode 100644 index 0000000..62ae132 --- /dev/null +++ b/results/figs/figure6_standard_depth_scaling_caption.md @@ -0,0 +1,18 @@ +# Standard-depth ResNet scaling figure caption + +**Figure 6 | Dynamic somato-dendritic innovation scales across standard +ResNet depth.** All points are frozen CIFAR-10 test endpoints from seeds +10--14 after 200 epochs; error bars show 95% normal intervals over paired +seeds. The renderer validates all 60 records and their source hashes against +the passed predeclared gate. **a,** BP, clean reciprocal Kolen--Pollack credit, +and dynamic SDIL under four-times-RMS soma-predictable apical traffic all gain +accuracy from ResNet-20 to ResNet-56. SDIL gains 1.176 points, and all five +paired seeds improve. **b,** On the full accuracy scale, fixed direct feedback +alignment remains far below both reciprocal local learners at every depth; +the complete nine-method crossover is reported separately rather than inferred +from this four-method panel. **c,** Paired ResNet-56 minus ResNet-20 gains for +BP and SDIL. **d,** SDIL retains a mean early-third teaching-signal cosine +above 0.9994 while its hardware-independent affine-MAC estimate remains below +1.34 times matched BP. Exact gradients are diagnostic-only and never used for +learning. The SDIL rule uses no task-loss queries but does use one explicitly +counted instruction-off neutral observation per ordinary example. -- cgit v1.2.3