blob: 51d63dc7ca3e302d515e49ec918b47fb0e795a4d (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
|
**Released physical drift contains a locally predictable state-dependent
component.** A--B, gate-voltage trajectories from the four small-network
`DriftTests` files released with Dillavou et al. (Zenodo 15692914). Colored
points are the retained pre-saturation samples within 30 mV of the published
single-task solution lines; faint curves show the complete post-initialization
traces. C--D, tangential drift velocity along each solution line. A constant
bias model (dotted) is compared with a per-edge local affine model
`B_i(G_i)=b_i+k_i(G_i-G_ref_i)` (solid). With the pre-saturation threshold
fixed at 50 mV before each trace's observed terminal gate plateau, the affine
model reduces forward time-block test RMSE to 0.21 and 0.54 of the constant
model in the two task pairs. The direction holds for terminal headrooms of 30,
50, 80 and 100 mV using unsmoothed centered-difference velocities. This
descriptive reanalysis establishes a measured state-dependent component
representable by the proposed local predictor. It does not demonstrate that
SDIL improves physical task learning; that is the prospective P1 test.
|
