diff options
| -rw-r--r-- | THREE_PART_EVIDENCE.md | 138 |
1 files changed, 138 insertions, 0 deletions
diff --git a/THREE_PART_EVIDENCE.md b/THREE_PART_EVIDENCE.md new file mode 100644 index 0000000..6673b0f --- /dev/null +++ b/THREE_PART_EVIDENCE.md @@ -0,0 +1,138 @@ +# Three-part paper evidence + +## One-sentence claim + +Somato-dendritic innovation is a local correction for predictable teaching- +channel error. It transfers across digital local learners, preserves coupled +learning as the number of imperfect components grows, and can be implemented +with local sampling and subtraction in a hardware-realistic CLLN. + +The method remains: + +```text +r = a - P(z) +Delta w = eta * r * eligibility +``` + +`P` is fitted from instruction-off local observations. It receives no task +gradient, device constant, downstream weight, or BP signal. + +## Part 1: additive digital correction + +The first part asks whether the same operation can be attached to an existing +learning rule without replacing that rule. + +| Digital learner | Task and repetitions | Clean | Raw imperfection | Static calibration | SDIL | Status | +|:--|:--|--:|--:|--:|--:|:--| +| Dual Propagation | CIFAR-10 miniCNN, 5 seeds | 82.86% | 9.40% | — | 82.92% | Passed frozen confirmation | +| Equilibrium Propagation | FashionMNIST ConvHopfield, 5 seeds | 76.26% | 31.38% | 67.90% | 74.52% | SDIL beats raw in every seed; strict gate failed because one seed favors calibration | +| Digital coupled learning | released ring tasks, size ladder pilot | reported by size | 37.5–62.5% error | 2.5–30.0% error | 0–7.5% error | Full confirmation running | +| Overclamped coupled learning | released ring tasks, size ladder pilot | 0–2.5% error | 25–50% error | — | 0–2.5% error | Clean-only time constant selected before biased endpoints | + +Sources: + +- `results/contrastive_bias/c1_gate.json` +- `results/ep_bias/c1_gate.json` +- `results/coupled_ladder/p1_imperfection_pilot.json` +- `results/coupled_ladder/p1_bias_baseline_pilot.json` +- `results/coupled_ladder/p1_overclamp_selected_pilot.json` + +The Dual Propagation result establishes strong transfer. The EP result shows +transfer with a visible remaining clean gap and seed variance. Coupled learning +adds a system where the local variables map directly to circuit measurements. + +## Part 2: scaling digital coupled learning + +The periodic grid ladder uses side lengths `4, 8, 12, 16, 24, 32`, or 32 to +2,048 learnable edges. Every method receives the same released task, tiled +initial gate field, and component draw. Size-dependent update exposure is +selected from ideal coupled learning only. + +Pilot final classification error: + +| Edges | Clean | Same-RMS noise | Raw | Static calibration | SDIL | +|--:|--:|--:|--:|--:|--:| +| 32 | 0.0% | 0.0% | 50.0% | 2.5% | 0.0% | +| 128 | 0.0% | 0.0% | 62.5% | 2.5% | 0.0% | +| 288 | 2.5% | 5.0% | 52.5% | 2.5% | 2.5% | +| 512 | 2.5% | 10.0% | 47.5% | 5.0% | 2.5% | +| 1,152 | 2.5% | 17.5% | 50.0% | 10.0% | 7.5% | +| 2,048 | 2.5% | 22.5% | 37.5% | 30.0% | 2.5% | + +The raw learner has a large error floor at every size. Static calibration works +on small grids and degrades on larger grids. Relative to static calibration, +SDIL reduces the pilot excess-error growth slope by 88.1%; the paired +task-bootstrap slope-difference interval is positive. At 2,048 edges, SDIL +closes the full raw-to-clean final-error gap and reaches stable zero error on +80% of task runs, versus 0% for raw and static calibration. + +The tuned overclamp pilot gives the same structural result. Clean overclamping +ends at 0–2.5% error across the ladder. Imperfect overclamping ends at 25–50%, +while overclamping plus SDIL ends at 0–2.5%. SDIL therefore combines with the +strong-clamp correction. + +The publication figure and data are: + +- `results/figs/figure_clln_scaling_pilot.pdf` +- `results/coupled_ladder/p1_scaling_analysis.json` +- `results/coupled_ladder/p1_scaling_source.csv` + +The pilot has five task clusters and one component draw per task and size. The +running confirmation contains all 40 released tasks, three new component draws, +six sizes, and five core methods: 3,600 training trajectories. Its outputs are +`results/coupled_ladder/p2_confirm_side{4,8,12,16,24,32}.json`. + +## Part 3: hardware-realistic simulation + +This part uses the nonlinear conductance law, periodic 4-by-4 topology, +released Figure-5 tasks and initial gates, Appendix-C component imperfections, +and explicit local voltage-square updates. + +Across 40 tasks and four component draws, or 160 trials per method: + +| Method | Mean classification error | Zero-error trials | +|:--|--:|--:| +| Clean CLLN | 0.00% | 100.00% | +| Raw imperfect CLLN | 25.86% | 21.25% | +| Static calibration | 3.75% | 84.38% | +| Overclamping | 4.77% | 86.88% | +| SDIL | 0.00% | 100.00% | +| Overclamping + SDIL | 0.00% | 100.00% | + +The local CDS/autozero circuit model then adds sampling gain mismatch, +pedestal mismatch, noise, and stale refresh. Ideal CDS and the combined mild +refresh-every-four condition retain 0% error. The combined strong condition +has 0.55% mean error and 96.88% zero-error trials. A 0.1 V/s pedestal mismatch +has 0.63% error; a 0.25 V/s mismatch has 4.45% error and marks the simulated +failure boundary. + +Sources: + +- `sdil/physical_grid.py` +- `results/physical_bias/p5_full_grid_bias_crossover.json` +- `results/physical_bias/p9_grid_correlated_autozero_key_results.json` +- `results/physical_bias/p8_spice_autozero_primitive.json` + +The hardware claim is: the full update is simulated with published device +equations and nonideal local sampling, and its primitive sample/subtract path +has a SPICE check. A fabricated-chip demonstration remains future work. + +## Main figures + +1. Method and transfer across Dual Propagation, EP, standard CLLN, and + overclamped CLLN. +2. Digital CLLN scaling: final error, stable success, and learning curves. +3. Hardware-realistic CLLN: raw, calibration, overclamping, SDIL, and + nonideal-CDS robustness. +4. Mechanism boundary: matched noise, state dependence, sampling mismatch, + and refresh interval. + +## Remaining gates + +1. Complete and audit the 3,600-trajectory digital ladder confirmation. +2. Run the selected overclamp pair on the full task/device confirmation panel. +3. Turn the local projection and scaling argument into a theorem whose plotted + quantity is excess task error or displaced fixed point, rather than an + unobserved residual alone. +4. Rebuild the three main figures from confirmation data and audit every number + against its source JSON. |