diff options
Diffstat (limited to 'slides/visual-contract.md')
| -rw-r--r-- | slides/visual-contract.md | 19 |
1 files changed, 9 insertions, 10 deletions
diff --git a/slides/visual-contract.md b/slides/visual-contract.md index 86a7ffc..26fd3d6 100644 --- a/slides/visual-contract.md +++ b/slides/visual-contract.md @@ -1,14 +1,13 @@ # SDIL project introduction deck: visual contract -- **Artifact:** Seven-slide, 16:9 project introduction deck. -- **Audience:** Machine-learning researchers with no prior SDIL context. -- **Core claim:** A local learner should use the soma-unpredicted component of a mixed apical signal; this residualization protects an inherited scalable credit path from predictable traffic. -- **Reader questions:** Why is raw feedback insufficient? What exactly is new? Is the update BP-free? Is residualization necessary? Does the combined method scale? Which evidence is controlled or synthetic? -- **Evidence layers:** problem and mechanism (slides 2–3), causal ablation (slide 4), standard-depth scaling and cost (slide 5), dynamical-task evidence (slide 6), attribution and boundary (slide 7). -- **Source data:** `results/figs/figure3_innovation.png`, `results/figs/figure5_bci_v2.png`, `results/oral_a_dynamic_scaling_v2_gate.json`, and the evidence-bound manuscript. -- **Statistics:** Five seeds for the controlled traffic and ResNet panels; six task clusters by five model seeds for the synthetic BCI panel. Values shown are audited means or paired outcomes already reported in the manuscript. -- **Visual grammar:** Direct mechanism diagrams, two result figures, and editable line charts. SDIL is blue, reciprocal KP is green, DFA is orange, BP is dark gray. +- **Artifact:** Seven-slide, 16:9 English project introduction deck. +- **Audience:** Machine-learning researchers new to SDIL. +- **Core claim:** A local learner can use the soma-unpredicted component of a mixed apical signal; residualization protects an inherited scalable credit path from predictable traffic. +- **Reader questions:** What problem does mixed feedback create? How does the learning rule work? Is residualization load-bearing? Does the combined system scale? Which component supplies each capability? +- **Evidence layers:** biological motivation and rule (slide 2), causal ablation (slide 3), standard-ResNet confirmation (slide 4), standard-depth scaling (slide 5), dynamical-task evidence (slide 6), attribution and next experiment (slide 7). +- **Source figures:** `results/figs/figure3_innovation.png`, `results/figs/figure4_resnet_confirmation.png`, `results/figs/figure5_bci_v2.png`, and `results/figs/figure6_standard_depth_scaling.png`. +- **Statistics:** Five seeds for the controlled traffic and ResNet panels; six task clusters by five model seeds for the synthetic BCI panel. +- **Visual grammar:** White background, black text, plain typography, full experimental figures, and cropped experimental panels. One equation slide carries the method. The deck adds zero process diagrams and zero redrawn result charts. - **Exact method labels:** SDIL, raw apical signal, somatic prediction, innovation, reciprocal KP, BP, DFA, local update, neutral observation. - **Output:** Editable PPTX, PDF export, generation source, and rendered QA contact sheet. -- **Claim boundary:** Clean scaling is attributed to reciprocal KP; SDIL-specific evidence is robustness under predictable mixed traffic and the synthetic BCI mechanism test. - +- **Attribution:** Reciprocal KP supplies clean-setting scale. SDIL supplies soma-conditioned traffic removal and the associated mechanism evidence. |