blob: 2de4e451b0a866e93956111e927694d98b1b21db (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
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.
|
