blob: 62ae132c1ea7f2e54fe5d606d524ac3a87bf5533 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
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.
|
