diff options
Diffstat (limited to 'paper/MANUSCRIPT.md')
| -rw-r--r-- | paper/MANUSCRIPT.md | 150 |
1 files changed, 125 insertions, 25 deletions
diff --git a/paper/MANUSCRIPT.md b/paper/MANUSCRIPT.md index b4b735a..733284e 100644 --- a/paper/MANUSCRIPT.md +++ b/paper/MANUSCRIPT.md @@ -29,13 +29,16 @@ dynamic paired-neutral innovation rule reaches 91.584% mean CIFAR-10 test accuracy across five untouched seeds, versus 91.388% for clean reciprocal credit; the one-sided 95% upper bound on its deficit is 0.131 points. It uses zero task-loss queries and 1.326 times the matched BP MAC estimate, but pays for -one instruction-off neutral observation per training example. A preregistered -synthetic BCI confirmation learns the task and causal sign yet fails outcome -vectorization and longitudinal prediction. The supported conclusion is -therefore algorithmic and narrow: neutral somato-dendritic innovation can -protect local credit from soma-predictable traffic and remain stable on -ResNet-20, without establishing a cortical learning rule or positive utility -from added standard-network depth. +one instruction-off neutral observation per training example. In a separate +six-task by five-model synthetic BCI confirmation, a local actor--critic +innovation reaches 100.000% task success, 99.833% terminal outcome decoding, +and 30/30 predicted causal-role signs. An acute outcome lesion reduces +role-aligned separation by 0.400. The supported conclusion remains +algorithmic: somato-dendritic innovation can protect local credit from +soma-predictable traffic, remain stable on ResNet-20, and multiplex performance +change with outcome surprise in a controlled dynamical task. These results do +not establish a cortical learning rule or positive utility from added +standard-network depth. ## 1. Introduction @@ -86,10 +89,9 @@ Our contributions are: 3. frozen evidence that the innovation operation is load-bearing under soma-predictable traffic, including an independently confirmed ResNet-20 endpoint; and -4. retained negative results showing where the proposal does not work: - arbitrary top-down traffic, an unstable static ResNet predictor, a failed - desired-velocity/online-control interpretation, and an untouched - population-signature confirmation that fails its joint gate. +4. an independently confirmed local actor--critic instantiation in which + learned causal roles vectorize performance change and outcome surprise, + together with retained failed protocols that delimit the result. ## 2. Somato-dendritic innovation learning @@ -196,6 +198,37 @@ loss, downstream weight, or reverse pass. It is nevertheless a real instruction-off microphase: every neutral observation and its elementwise arithmetic are counted, and we do not call this variant single-phase. +### 2.4 Temporal-difference outcome innovation + +The synthetic BCI branch uses the same innovation variable in a continuous +dynamical task. A local linear critic reads the surrounding somatic population, + +\[ +V_t=v^\mathsf{T}[1,h_t], +\qquad +\delta_t=\rho_t+\gamma(1-z_t)V_{t+1}-V_t, +\] + +where \(z_t\) marks a terminal event and +\(\rho_t=|e_{t-1}|-|e_t|+\mathbb{1}\{\text{rewarded terminal}\}\). +Sparse antithetic cursor probes estimate each cell's signed causal role +\(m_i=\partial z/\partial h_i\); the instructional apical term is +\(m_i\delta_t\). A local temporal eligibility trace, + +\[ +E_{ij,t}=0.8E_{ij,t-1}+ +(1-h_{i,t}^2)x_{j,t}, +\qquad +\Delta W_{ij,t}=\eta r_{i,t}E_{ij,t}, +\] + +assigns the innovation to recent synaptic activity. Actor, critic, role +estimator, and neutral predictor use manual local updates without autograd or +task-loss queries. The terminal reward is explicitly supplied, so outcome +encoding alone is not evidence for an emergent error code; the causal tests +are residualization, learned role vectorization, critic expectedness, acute +lesions, and untouched replication. + ## 3. What residualization guarantees—and what it does not Let \(n\) be neutral apical traffic and let @@ -279,6 +312,25 @@ and dynamic innovation from scratch for untouched seeds 10--14, uses no validation examples, and evaluates the 10,000-example test set once at the endpoint. Both conditions use the same initialization/data seed pairing. +### Synthetic BCI confirmation + +The biological-signature test uses 40 cells with known experimenter-only +causal roles, 14 training days, 64 episodes per day, and 28 maximum steps per +episode. Every condition receives the same instruction-off neutral warmup and +paired scalar cursor probes. Fixed-role, plasticity-lesion, critic-lesion, +outcome-lesion, and exact-role diagnostic conditions share all exogenous +trajectories. + +Because independently learned policies have different cursor scales, a fixed +absolute target produced a disclosed ceiling failure. The final assay freezes +the algorithm, runs 512 separate outcome-free calibration trajectories, and +uses fixed cursor-maximum quantiles +\(\{0.20,0.35,0.50,0.65,0.80\}\) as target levels. Rewarded and timeout +outcomes are then measured on independent trajectories; no evaluation label +selects or reweights a target. Confirmation crosses six untouched task seeds +and five model seeds. Uncertainty first averages models within task and then +uses the six task seeds as independent clusters. + ### Baselines and cost In-repository controls include BP, FA, DFA, direct node perturbation, @@ -371,7 +423,7 @@ time is 1.47 times paired clean KP. The result supports ResNet-20 robustness under the audited traffic intervention. It does not support a ResNet-20-to-56 depth claim. -## 6. Biological-signature test and negative evidence +## 6. Biological-signature test and outcome-surprise evidence The original online-control/desired-velocity screen fails its frozen causal sign and acute-control gates. We therefore constructed a separate @@ -395,14 +447,59 @@ the broader population outcome-vectorization and longitudinal signatures. Because online control is disabled in this recovery, even a passed plasticity gate would not establish an online desired-velocity controller. +That failure exposed two structural mismatches rather than motivating a +threshold repair. First, the old trajectory terminated without an explicit +reward/timeout event, so the residual was never asked to encode the variable +used by its outcome decoder. Second, near-ceiling task success left almost no +unrewarded trials. A separately frozen v2 added an explicit actor--critic +outcome phase. Its first development grid failed from a cold start when dense +performance velocity was scaled to one quarter of the previously validated +signal. Restoring unit scale recovered learning, but a fixed target ladder +failed one new development seed because policy output scale varied. Both +failures remain in the repository. + +The final recovery changes no learning parameter after that diagnosis. It uses +the label-free calibration split described above and then opens a fully +untouched 30-record confirmation. + + + +**Figure 5: Role-vectorized temporal-difference outcome surprise.** The +renderer reads all 30 confirmation records and uses the task seed, not each +model, as the uncertainty unit. + +All six task clusters reach 100.000% final success; mean learning gain is +98.802 points and the fixed-role gap is 99.974 points. Learned-role cosine is +0.9761. Nonterminal residual--soma correlation is 0.063, while subtracting the +neutral prediction reduces absolute correlation by 0.936. The preceding +surrounding population predicts causal-cell residual sign at 54.37% balanced +accuracy, with a one-sided lower bound of 54.26%, and velocity has a 0.638 +absolute cross-validated correlation advantage over error magnitude. All 30 +records have the predicted positive P+/P- sign. + +Independent calibrated trials are balanced at 50.094% rewarded outcomes. +Terminal residual outcome accuracy is 99.833% with a 99.680% lower bound, +versus 73.50% from pre-outcome soma. Acute removal of terminal outcome input +reduces role-aligned separation by 0.400 (lower bound 0.389). The learned +critic contributes 0.319 of expectedness modulation (lower bound 0.281), and +that contribution is paired with its stored value prediction. + +The result establishes the claimed signal within this synthetic paradigm: +ordinary soma coupling is subtracted, recent performance change and terminal +outcome are vectorized by learned cell-specific causal roles, the residual +drives local eligibility-based plasticity, and the critic converts raw outcome +into surprise. It does not establish that the same plasticity rule operates in +cortex. Calibration makes rewarded and timeout trials statistically +identifiable; it is an explicit psychometric phase and not a biological +prediction by itself. + These negatives constrain interpretation: - we do not infer that dendritic residuals directly drive cortical plasticity; - we do not infer that cortex implements BP; - we do not claim arbitrary top-down traffic removal; - we do not claim positive utility from adding standard ResNet depth; and -- we do not relabel task learning and sign inversion as a passed population - signature. +- we do not treat directly supplied terminal reward as an emergent error. ## 7. Related work @@ -462,17 +559,19 @@ MAC overhead. Hardware implementations may price local elementwise operations, state storage, and phases differently from GPUs; this is why we report several resource axes rather than one scalar cost. -The strongest standard result uses only ResNet-20. The separately frozen -ResNet-20/32/56 panel remains unopened because its biological prerequisite -failed. This preserves the declared accept-to-oral ordering but leaves positive -added-depth utility unresolved. +The strongest standard result uses only ResNet-20. The old separately frozen +ResNet-20/32/56 panel remains unopened because its original biological +prerequisite failed. The successful v2 gate permits only a new independently +frozen depth protocol; it cannot retroactively open the old panel. Positive +added-depth utility therefore remains unresolved. -Finally, the synthetic BCI negatives are scientifically important. The -residual operation can be algorithmically useful without reproducing every -signature of cortical dendrites. A stronger biological paper needs a new -prediction and mechanism frozen independently of the failed outcome and -longitudinal metrics, ideally tested on the original event-level data rather -than engineered to pass the present synthetic task. +Finally, the synthetic BCI evidence has a hard ecological boundary. Outcome +reward is supplied to the critic, the psychometric targets are calibrated to +each trained policy on a separate split, and the task has experimenter-defined +causal roles. The acute lesions show how the implemented signal is composed; +they do not show that cortex uses the same decomposition. A stronger biological +paper needs a prospective prediction tested on the original event-level data, +and the still-failed longitudinal prediction should not be silently discarded. ## 9. Reproducibility statement @@ -486,4 +585,5 @@ bash experiments/finalize_accept.sh It rechecks the main figures, theoretical identities, local-rule mechanics, baseline protocols, native-author records, standard-ResNet confirmation, -failed biological confirmation, and the sealed standard-depth boundary. +failed biological protocols, the passed calibrated BCI confirmation, and the +sealed old standard-depth boundary. |