blob: 41e0dbdcb58e44a812d35c867b5f624979c1c465 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
|
# Oral-B-v2 confirmation figure caption
**Figure 5 | Role-vectorized somato-dendritic innovation carries
temporal-difference outcome surprise.** All summaries use the complete
untouched six-task by five-model confirmation; uncertainty is computed after
averaging models within each task seed. **a,** Mean episode success over
training. The learned causal-role vector is necessary, the forward-plasticity
lesion (omitted for visual overlap with zero) does not learn, and critic or
terminal-outcome training lesions are reported rather than hidden. Bands are
95% normal intervals over six task clusters. **b,** Ordinary apical traffic is
almost perfectly soma-correlated before subtraction; the innovation is below
the frozen correlation ceiling in every task cluster. **c,** Five targets are
chosen from fixed quantiles of 512 outcome-free cursor-max calibration trials.
The displayed rewarded fractions come from independent assay trajectories;
no evaluation label selects or reweights a target. **d,** Terminal innovation
decodes rewarded versus timeout outcomes, while the acute terminal-outcome
lesion removes role-aligned separation. The critic contribution tracks the
stored value prediction and attenuates expected outcomes. Exact role maps,
outcome labels, and gradients are diagnostic-only and never used by the
learner's updates.
|
