diff options
Diffstat (limited to 'results/figs/physical_bias_state_dependence_caption.md')
| -rw-r--r-- | results/figs/physical_bias_state_dependence_caption.md | 15 |
1 files changed, 15 insertions, 0 deletions
diff --git a/results/figs/physical_bias_state_dependence_caption.md b/results/figs/physical_bias_state_dependence_caption.md new file mode 100644 index 0000000..51d63dc --- /dev/null +++ b/results/figs/physical_bias_state_dependence_caption.md @@ -0,0 +1,15 @@ +**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. |