diff options
| -rw-r--r-- | NOVELTY.md | 106 | ||||
| -rw-r--r-- | RESULTS.md | 14 | ||||
| -rw-r--r-- | ROADMAP.md | 16 | ||||
| -rw-r--r-- | sdil/core.py | 24 |
4 files changed, 139 insertions, 21 deletions
diff --git a/NOVELTY.md b/NOVELTY.md new file mode 100644 index 0000000..3755595 --- /dev/null +++ b/NOVELTY.md @@ -0,0 +1,106 @@ +# SDIL novelty audit + +This file records the closest-prior-art boundary as of 2026-07-22. It is a claim constraint, not +a literature-survey completeness claim. New evidence may narrow the allowed claims further. + +## Exact overlap with learned node-perturbation feedback + +Lansdell, Prakash, and Kording, [Learning to solve the credit assignment +problem](https://arxiv.org/abs/1906.00889) (ICLR 2020), already train synthetic feedback matrices +from node-perturbation estimates. Their direct-feedback form maps the output error to every hidden +layer and was evaluated in a CIFAR convolutional network. Their public implementation is +[`benlansdell/synthfeedback`](https://github.com/benlansdell/synthfeedback). + +With no ordinary apical traffic and `P=0`, SDIL reduces exactly to this idea (up to sign and tensor +conventions): + +```text +c = output error +r_l = A_l c +q_l = node-perturbation estimate of -dL/dh_l +A_l <- A_l + eta_A (q_l - A_l c) c^T +``` + +Independent noise at multiple hidden layers during a shared noisy forward pass is also present in +the earlier method. SDIL's antithetic Rademacher directions, vectorized direction batching, and +explicit cost accounting are useful estimator/implementation choices, but are not the core +algorithmic novelty. The existing no-traffic depth results therefore establish the strength of a +learned-node-perturbation backbone; by themselves they do not establish a Harnett-specific +contribution. + +The closest exact baseline is consequently the same implementation with `traffic_mode=none`, +`learn_P=0`, and `use_residual=0`. It should be named **learned direct feedback by node +perturbation (Lansdell et al.)**, not SDIL, in comparisons intended to isolate novelty. + +## Overlap with dendritic credit assignment + +Apical/basal segregation, neuron-specific dendritic teaching signals, and local three-factor or +burst-dependent plasticity substantially predate SDIL. In addition, Greedy et al., +[Cell-type-specific cortical feedback coordinates hierarchical credit +assignment](https://www.biorxiv.org/content/10.64898/2026.06.16.732595v1) (BurstCCN, 2026), +explicitly connects a scalable dendritic credit-assignment model to the Francioni/Harnett data. +BurstCCN already: + +- models deviations from balanced dendritic feedback as neuron-specific error signals; +- reproduces the opposite residual signs of the P+ and P- BCI populations and their disruption by + NDNF activation; +- trains convolutional models on CIFAR-10 and ImageNet with plastic feedback; and +- provides an [author implementation](https://github.com/neuralml/BurstCCN-journal). + +Thus neither "apical dendrites carry vector errors", "the model reproduces Harnett Fig. 5", nor +"a dendritic local rule scales to ImageNet" is a sufficient novelty claim. + +## Candidate SDIL contribution + +The candidate contribution is the combination + +```text +a_l = A_l c + t_l mixed apical traffic +P_l(h_l) ~= E[t_l | h_l, neutral] per-cell neutral predictor +r_l = a_l - P_l(h_l) somato-dendritic innovation +Delta W_l proportional to r_l * eligibility_l +``` + +where the subtractive per-cell innovation, rather than raw apical activity, is the teaching +variable. The feedback vectorizer `A_l` may be learned using Lansdell-style causal perturbations; +that component is inherited, not claimed as new. The new empirical question is whether +residualization is load-bearing when teaching signals share an apical channel with ordinary +somatic/contextual feedback. + +Neutral-period fitting is an algorithmic adaptation rather than a direct statement of the Harnett +paper. It addresses an identifiability problem: task-period regression can subtract the predictable +part of `A_l c` whenever output error correlates with somatic state, while neutral-period regression +identifies the traffic relationship without observing the teaching component. + +## Claims currently allowed + +- The inherited learned-feedback backbone preserves performance with depth better than fixed DFA + in the audited flattened-CIFAR MLP setting. +- Controlled pilots show that per-cell residualization can protect learning from soma-coupled and + endogenous top-down traffic; the latter remains a validation-only, single-seed result until C1 is + completed. +- Simultaneous batched perturbations make the inherited calibration practical in this codebase; + a hardware-independent query/FLOP advantage is not yet established. + +## Claims currently forbidden + +- Node perturbation training of `A_l` is novel. +- Simultaneously perturbing all hidden layers is novel. +- The no-traffic scaling panel demonstrates the necessity of somato-dendritic residuals. +- SDIL is the first dendritic method to scale to deep vision tasks. +- Reproducing the already-published Harnett/BurstCCN qualitative signatures is an oral-level + biological contribution. + +## Evidence that can restore a strong paper + +1. Show across independently generated, endogenous traffic families that residualization beats + raw and norm-matched raw feedback, while the no-traffic methods coincide. +2. Compare directly with Lansdell-style learned feedback and BurstCCN under validation-selected, + disclosed compute/query budgets. +3. Prove the conditional-bias/SNR benefit of neutral residualization and test its predicted failure + when traffic depends on the teaching signal after conditioning on soma. +4. For a biological contribution, make and test a prediction not used by Harnett or BurstCCN. Raw + event-level data and analysis code require author coordination; the paper exposes only plotted + source data and states that full data/code are available on request. +5. For a scaling contribution, target standard architectures and a simplicity/cost regime not + already occupied by BurstCCN, rather than treating ImageNet alone as novelty. @@ -32,7 +32,8 @@ Per hidden layer l: - apical feedback `a_l = A_l c` (c = output error e = softmax−onehot, broadcast) - default per-neuron predictor `ĥ_l = p_l ⊙ h_l + b_l`; teaching signal `r_l = a_l − ĥ_l` - three-factor update `ΔW_l = η (r_l ⊙ φ'(u_l)) h_{l-1}^T` -- `A_l` LEARNED by amortized node perturbation (q_l ≈ −∇_{h_l}L, forward-only, no weight transport) +- `A_l` learned by node perturbation (q_l ≈ −∇_{h_l}L, forward-only, no weight transport), + following Lansdell, Prakash & Kording's learned synthetic-feedback method - `P_l` learned on neutral (c=0) periods only, so it strips the soma-predictable nuisance without eating the teaching signal. @@ -40,9 +41,11 @@ The scalable perturbation mode injects independent Rademacher interventions at a in the same antithetic forward evaluations. Cross-layer interference adds variance but averages to zero; it trains all `A_l` using `O(depth)` rather than `O(depth^2)` forward work. -Distinct from the failed `~/sdrn` (fixed-random feedback + post-hoc residualization) and from -Dual Prop / DFA / FA: the feedback pathway is *learned by causal perturbation*, so the residual is -aligned by construction rather than being a random projection. +Relative to fixed DFA/FA, the feedback pathway is learned by causal perturbation. That ingredient +is not novel: without traffic and `P`, it is the direct learned-feedback method of +[Lansdell et al.](https://arxiv.org/abs/1906.00889). The candidate SDIL contribution is specifically +the per-cell residual under mixed apical traffic and its neutral-period identification. The full +claim boundary is frozen in `NOVELTY.md`. ## Critical fix Node-perturbation estimator must divide by σ, not σ² (with δ=σξ, ĝ = δ·ΔL/σ² = ξ·ΔL/σ ⇒ ‖q‖≈‖∇L‖). @@ -212,6 +215,9 @@ their captions, source-hash manifest, and `results/audited_tables.md`, then runs ## Open items +- Treat learned direct node-perturbation feedback (Lansdell et al.) as the exact backbone baseline, + and BurstCCN as the strongest dendritic/scaling baseline; do not attribute their ingredients to + SDIL. - Compare simultaneous calibration at fixed loss-evaluation budgets and sweep directions (4/8/16/32). - Validate SDIL on a genuinely depth-necessary compositional task and a CNN; flattened CIFAR does @@ -59,8 +59,11 @@ must hold in a hardware-independent cost coordinate as well as GTX-1080 wall tim ### C4. The comparison set contains the closest alternatives -The main comparison must include exact BP, FA, DFA, direct/unamortized node perturbation, Dual Prop, -and a burst/dendritic credit-assignment method such as BurstCCN. EP remains an informative +The main comparison must include exact BP, FA, DFA, direct/unamortized node perturbation, learned +direct node-perturbation feedback (Lansdell et al. 2020), Dual Prop, and BurstCCN. The Lansdell +method is the exact no-traffic/P=0 backbone of the current implementation, not merely a related +baseline. BurstCCN is already demonstrated on CIFAR-10 and ImageNet and already models the Harnett +BCI signatures, so it is the strongest dendritic/scaling comparator. EP remains an informative relaxation-based baseline. Forward-Forward and PEPITA are appendix context unless they become competitive under the frozen protocol. @@ -108,10 +111,11 @@ claim is explicitly computational rather than a fit to cortical data. ## Oral bar A: standard deep architectures Prepare convolutional local-update primitives and ResNet-20/32/56 protocols early. Queue frozen -runs opportunistically on authorized idle GPUs. The oral-level target is a memorable joint result: -near-BP accuracy in a genuinely deep standard model, a win over the strongest local baseline, and -a nondominated accuracy–hardware-independent-cost point. Scale alone is not enough if C1 does not -show that somato-dendritic innovation is the load-bearing operation. +runs opportunistically on authorized idle GPUs. Because BurstCCN already reports CIFAR-10 and +ImageNet scaling, dataset scale alone is not novel. The oral-level target is a memorable joint +result: near-BP accuracy in a standard architecture, a win over or materially simpler/lower-cost +regime than BurstCCN, and a nondominated accuracy–hardware-independent-cost point. Scale alone is +also insufficient if C1 does not show that somato-dendritic innovation is load-bearing. ## Stop conditions diff --git a/sdil/core.py b/sdil/core.py index 37c7165..cece194 100644 --- a/sdil/core.py +++ b/sdil/core.py @@ -4,11 +4,11 @@ SDIL -- Somato-Dendritic Innovation Learning. A non-backprop learning rule inspired by Harnett et al. 2026 (Nature), "Vectorized instructive signals in cortical dendrites". -The load-bearing idea, faithful to the paper, is NOT "apical dendrites carry -error". That is old (Guerguiev/Sacramento/Payeur/burst credit assignment). The -new algorithmic object is the *somato-dendritic residual*: the teaching signal -is the apical feedback MINUS the part of it that is predictable from the cell's -own somatic activity. Only the innovation drives plasticity. +The Harnett-specific object studied here is the *somato-dendritic residual*: +the teaching signal is the apical feedback MINUS the part of it that is +predictable from the cell's own somatic activity. Only the innovation drives +plasticity. Apical credit signals and learned feedback from node perturbations +both have substantial prior art; see NOVELTY.md for the exact claim boundary. For hidden layer l: u_l = W_l h_{l-1}, h_l = phi(u_l) (somatic / basal forward) @@ -21,10 +21,11 @@ Three learning rules, all local (no weight transport, no reverse-mode graph): dP_l = eta_P * (a_l - P_l h_l) h_l^T (slow; only on neutral periods) eta_P << eta dA_l = eta_A * (q_l - r_l) c_l^T (on perturbation trials; q_l = causal node-pert estimate) -q_l is an amortized node-perturbation estimate of the causal descent direction +q_l is a node-perturbation estimate of the causal descent direction -grad_{h_l} L, obtained by occasionally perturbing h_l and watching the loss. -Learning A_l this way (rather than fixing it random) is what keeps SDIL from -degenerating into DFA. +Learning A_l this way follows the learned synthetic-feedback approach of +Lansdell, Prakash & Kording (ICLR 2020). Relative to fixed DFA it supplies an +adaptive feedback backbone; it is not itself claimed as an SDIL novelty. This module deliberately does NOT use autograd for learning. Autograd is used ONLY inside probes.py to MEASURE the true gradient for alignment diagnostics. @@ -92,9 +93,10 @@ class SDILNet: biologically unproblematic; only hidden layers use the SDIL innovation. feedback signal c: by default the output error e = softmax(logits) - onehot, - broadcast to every hidden layer (dim = n_classes). With A_l learned this is - "amortized node perturbation through apical dendrites"; with A_l fixed random - it reduces to DFA (used as an ablation). + broadcast to every hidden layer (dim = n_classes). With P=0 and no traffic, + learning A_l from node perturbations is the direct learned-feedback method + of Lansdell et al.; with A_l fixed random it reduces to DFA. SDIL's candidate + addition is the per-cell innovation under mixed apical traffic. """ def __init__(self, sizes, act="tanh", device="cpu", seed=0, |