# Somato-Dendritic Innovation Learning (SDIL) SDIL is a non-backpropagation learning project motivated by the neuron-specific somato-dendritic residuals reported by Harnett and colleagues. Its central algorithmic hypothesis is narrow: a mixed apical compartment should teach from the component that is unexpected given the same neuron's ordinary somatic state, rather than from raw apical activity. For hidden population `l`, the implemented rule is ```text a_l = A_l c + ordinary apical traffic a_hat_l = P_l(h_l), fitted during neutral periods r_l = a_l - a_hat_l Delta W_l = eta (r_l * local postsynaptic gain) h_(l-1)^T. ``` The standard-ResNet stability branch additionally uses a fast paired neutral observation to project any *remaining* affine neutral residual off current soma before plasticity. This local microphase is explicitly counted; it is not described as single-phase or free. `A_l` is calibrated from antithetic node perturbations without reverse-mode differentiation or weight transport. That learned-feedback mechanism is inherited from Lansdell, Prakash, and Kording (2020); it is not claimed as novel. The candidate contribution is the neutral-period, per-cell innovation operation under mixed apical traffic, together with its theory, causal audit, and scaling behavior. See `NOVELTY.md` for the exact prior-art boundary. ## Current audited evidence - On flattened CIFAR-10, SDIL loses only `0.214 +/- 0.349` accuracy points from hidden depth 5 to 60, while DFA's early-layer teaching alignment falls from `0.514` to `0.047`. This is depth preservation on a depth-flat task, not yet evidence that added depth is useful. - With strong soma-predictable apical traffic, raw and norm-matched raw learning fall to `10.38%` and `10.31%`, while innovation learning retains `97.35%`. The paired innovation gain over norm-matched raw is `87.038 +/- 0.607` points across five seeds. - Direct, unamortized node perturbation passes the controlled useful-depth diagnosis, but costs `68.4x` ordinary forward-equivalent work. The learned apical vectorizer, rather than the local eligibility rule, is the current useful-depth bottleneck. - A low-query K1/every-4 calibration protocol retains `112.9%` of the K16 gain over DFA with 16x fewer logical queries, 11x less calibration work, and 5.3x less total forward-equivalent work in the frozen MLP protocol. - The frozen standard-ResNet funnel passed its BP reference and short screen: BP reached `91.62%`, and short-run channel-gated SDIL reached `41.98%` versus tuned DFA's `37.16%`. The full A3 run then failed decisively: SDIL became nonfinite at epoch 89 and ended at chance, while DFA remained finite at `33.06%`. A4 was therefore not opened and no confirmation test seed was touched. - The post-failure representable-subspace estimator passed its exact mechanics tests and improved frozen-forward early/all-layer alignment from `0.0011/0.0117` to `0.0072/0.0527` at matched query count. It nevertheless missed the frozen `0.01` early-layer and improvement gates, so no v2 full ResNet run was launched. - Direct vectorizer-space perturbation then passed its exact mechanics and synthetic variance audit, but V3-1 again failed: early alignment was `0.00714` versus matched V2's `0.00721`, although all-layer alignment rose to `0.06258`. V3 full training therefore remained closed. - Residual response mirroring passes its short gate but fails its sole full run at 10% with NaN validation loss. Its endpoint Q/W cosine nevertheless reaches 0.999998, exposing final alignment as an inadequate certificate for an intermittent tracker. The separate reciprocal KP short gate reaches 82.66%, and its frozen full gate reaches 91.26% versus matched BP's 91.62% with 0.9997 late feedback cosine. The subsequent frozen mixed-traffic screen nevertheless makes raw, norm-matched raw, and innovation all nonfinite in epoch 1. A separately frozen operator-stability branch then rules out fixed predictor margins and passes a dynamic paired-neutral controller: its D1 352-step trajectory is finite, and D2 reaches `83.58%` after 20 epochs versus clean KP's `82.66%` and the failed mixed-traffic controls' `10%`, at `1.3261x` BP MACs and zero task-loss queries. This is still one short validation seed; the predeclared 200-epoch D3 run subsequently passes all 19 gates at `91.18%` versus BP's `91.62%` and clean KP's `91.26%`, with `0.9994` early alignment, zero task-loss queries, and `1.3261x` BP MACs. This raises the strict score to 6/10. The independently frozen paired five-seed D4 test panel then passes every gate: dynamic innovation reaches `91.584%` mean test accuracy versus clean KP's `91.388%`, wins the mean pairing by `0.196` points, and has a clean-minus-dynamic one-sided 95% upper bound of only `0.131` points. Every dynamic seed reaches at least `91.51%`, mean early alignment is `0.999687`, and all trajectory, projection, leakage, query, hardware, MAC, memory, split, and test-isolation checks pass. This untouched confirmation establishes the strict 7/10 accept bar; it does not establish that added standard-network depth is useful or repair the failed biological signatures. - Native author-code fidelity is complete. BurstCCN reaches `80.10%` at its validation-selected epoch versus published `82.97 +/- 0.21%`; Dual Prop reaches `92.46%` versus published `92.41 +/- 0.07%`. Their audited walls are `15451.3 s` and `23119.8 s`, respectively, and they are not presented as matched-compute points. The broad endogenous-traffic gate and the Harnett desired-velocity/online-control screen failed. Those results are retained and explicitly constrain the paper: SDIL does not currently explain the reported error-derivative signature, and the residual mechanism is supported for traffic predictable from the chosen somatic statistic rather than arbitrary top-down context. `ROADMAP.md` and `ORAL_B.md` contain the frozen gates and complete negative branches. A mechanics-only recovery now separates perturbation-learned causal role from within-episode performance velocity; it passes its algebraic checks but has no task endpoint and does not yet weaken this limitation. ## Publication-facing artifacts - `RESULTS.md`: audited positive and negative results; - `THEORY.md`: estimator variance, descent conditions, conditional innovation, timescales, and hardware-independent cost; - `BASELINES.md`: matched and native-author baseline ledger; - `ROADMAP.md`: accept/oral evidence gates and their state; - `MIXED_TRAFFIC.md`: frozen standard-ResNet raw/matched/innovation accept gate; - `DYNAMIC_INNOVATION.md`: post-failure two-timescale stability theory and frozen D1--D4 gates; - `ORAL_B_RECOVERY.md`: structural diagnosis and mechanics-only temporal- difference recovery boundary; - `ORAL_A.md`: frozen standard CIFAR ResNet funnel; - `ORAL_A_V2.md`: frozen post-failure representable-subspace funnel; - `ORAL_A_V3.md`: frozen vectorizer-space causal-calibration funnel; - `REVIEW_SCORECARD.md`: adversarial ICLR-style score trajectory; - `results/figs/`: deterministic PDF/PNG main figures, captions, and a source hash manifest, plus audited RRM failure and dynamic-stability supplements. The three current main figures show the local-method Pareto frontier, credit assignment versus depth, and the load-bearing innovation ablation. Every final cell has exactly seeds 0--4 and clean git provenance; the strict renderer refuses missing or protocol-mixed cells. ## Verification The lab environment currently used for audited CPU checks is `/home/yurenh2/miniconda3/envs/ep_pascal`. Run: ```bash experiments/finalize_claims.sh ``` This regenerates the figures and manifest, enforces the frozen Pareto/scaling gates, verifies baseline and protocol mechanics, checks the local-rule smoke tests, evaluates the theory identities, and rebuilds the audited tables. The convolutional infrastructure can be checked independently: ```bash /home/yurenh2/miniconda3/envs/ep_pascal/bin/python3 experiments/cifar_image_smoke.py /home/yurenh2/miniconda3/envs/ep_pascal/bin/python3 experiments/conv_local_smoke.py ``` These checks cover deterministic CIFAR splitting/augmentation, the standard CIFAR `6n+2` ResNet topology, exact local convolution and BatchNorm eligibilities, the BatchNorm-coupled perturbation objective, predictor fitting, translation-shared feedback, representable-subspace causal calibration, and parameter/cost accounting. The latter is a transparent post-A3 development path: it perturbs the channel-gated basis directly and reproduces the expected full-field delta rule with far less irrelevant spatial variance. It does not retroactively reopen or relabel the failed frozen Oral-A protocol. The same smoke suite also audits a convolutional hierarchical-FA baseline. Its random feedback tensors follow the real residual DAG and use local activation Jacobians, but remain independent of forward weights. An audit-only symmetric copy reproduces exact hidden gradients to `5.3e-16` relative error and the BP parameter update to `2.98e-8`; actual HFA never performs that copy. `HFA_BASELINE.md` freezes its matched ResNet-20 validation screen before any HFA accuracy endpoint is observed. The learned-hierarchy development path perturbs the feedback convolution parameter spaces themselves. One antithetic pair estimates every edge's causal target moment simultaneously, while the current prediction moment is a local child/parent correlation. Its BatchNorm-coupled JVP and symmetric-limit delta rule are executable smoke tests. `ORAL_A_V4.md` freezes the causal-capture and conditional accuracy gates; no accuracy claim is attached until they pass. Because the global-scalar V4 gate failed, the next strong inherited comparator uses normalized local response mirroring. Its observation phase sends Gaussian parent activity through ordinary local forward synapses; a separate feedback update sees only the probe/child-response pairs. The smoke suite verifies both high finite-sample kernel recovery and that changing forward parameters after observation cannot affect the update. Any scale obtained this way is credited to the weight-mirror baseline until innovation residualization is shown to be load-bearing on top of it. `MIRROR_BASELINE.md` freezes that comparator's CIFAR-prefix capture, short accuracy, and conditional full-validation gates. After estimate-then-average WM narrowly failed its short accuracy gate, the next substantive inherited control uses residual response mirroring. Q predicts the observed child response and learns only from the local prediction error; unlike WM, its update is zero for every probe at exact symmetry. The extra feedback prediction convolution is charged explicitly, and this rule receives no SDIL novelty credit. `RESIDUAL_MIRROR.md` freezes its capture and conditional task gates. The stronger reciprocal Kolen--Pollack substrate is audited separately in `KP_BASELINE.md`. Its short and full frozen gates pass; the full record reaches 91.26% validation accuracy, 0.9994 early alignment, and 1.326x BP MACs without feedback loss queries. Before that full endpoint, `MIXED_TRAFFIC.md` froze the actual Harnett-specific test: raw, norm-matched raw, and innovation under identical four-times-RMS soma-predictable apical traffic and predictor cost. Mechanics pass, but the complete MT-1 panel fails: all three signals become nonfinite in epoch 1 and end at 10.00% validation accuracy. MT-2 and MT-3 remain untouched, closing this standard-ResNet recovery without a weaker traffic intervention. The subsequent V3 mechanism estimates the required A/G matrix statistics directly by perturbing the vectorizer parameter subspace. It remains forward-only and uses two causal loss queries per direction. Its exact and BatchNorm JVP identities, delta rule, and matched synthetic variance comparison are covered by `experiments/conv_local_smoke.py`; no task-level claim is made until a new frozen gate is completed. `experiments/finalize_accept.sh` additionally requires strict imports of the frozen BurstCCN and Dual Propagation author-code runs. Both records now pass; the accept-bar mechanical audit is green. The standard ResNet branch stopped at its failed A3 validation gate, and the untouched A4 test panel remains sealed. ## Result discipline Every publication-facing run records the source revision, tracked-tree state, full arguments, split identity, seed, evaluation count, wall time, query/MAC accounting, and final finiteness. Development, validation, and untouched confirmation results are never pooled. Failed gates close their branch instead of triggering seed deletion or post-hoc threshold changes. The current strict reviewer estimate is `5/10` (borderline reject, confidence `4/5`): the mechanism and controlled depth-preservation results are strong, but the frozen standard-useful-scale attempt failed. The score changes only after an audited frozen stage, not after a pilot or a presentation improvement.