1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
|
# 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.
Likewise, recursive spatial feedback, learned feedback convolutions, and local
ReLU-gated error propagation are not new by themselves. The post-failure
hierarchical oracle deliberately reconstructs a BP-like local adjoint and may
guide engineering, but it cannot be presented as an SDIL contribution. Any
trainable version must be compared directly with hierarchical FA, learned
feedback alignment/weight-mirroring methods, and BurstCCN. The Harnett-specific
claim remains the neutral-period somato-dendritic innovation operation under
mixed traffic.
The implemented feedback-parameter perturbation rule is task-causal rather
than a copied-weight or forward-reconstruction rule, but this distinction does
not by itself make recursive learned feedback novel. It is treated as a
credit-path engineering component unless the complete innovation model shows a
load-bearing residualization result beyond those mandatory comparators.
Likewise, the normalized local response-mirror path is explicitly inherited
from weight mirroring/weight estimation. Even if it recovers BP-level scale,
that performance belongs to the baseline. A paper-level SDIL gain requires the
somato-dendritic innovation subtraction to remain necessary on top of the
mirrored hierarchical path under a frozen mixed-traffic intervention.
Residual response mirroring changes the inherited baseline's estimator but not
this attribution. Its zero-noise fixed point and any resulting scale belong to
local predictive weight estimation; they become relevant to SDIL only as the
feedback substrate on which innovation is separately ablated.
The modified Kolen--Pollack reciprocal-plasticity path is likewise inherited
from Akrout et al. Equal local activity products and matched decay in reciprocal
synapses are not SDIL. Even if the audited implementation reaches BP-level
accuracy, it only supplies a stable feedback substrate. The candidate novelty
still requires neutral somato-dendritic subtraction to outperform raw and
norm-matched raw apical activity under the same mixed-traffic intervention.
The post-MT-1 fast neutral projection does not make reciprocal KP novel. It
adds a second-timescale operation to the innovation mechanism:
```text
e0_l = a0_l - P_l(h_l) neutral residual
q_l = Cov(e0_l,h_l) / Var(h_l) current local coupling
r_l = s_l + e0_l - E[e0_l] - q_l(h_l-E[h_l]) stabilized innovation
```
The coefficient fit is local and instruction-off, so it does not use a task
loss, task-nudged state, downstream weight, or reverse pass. Its defensible
novelty is not “two phases” or generic normalization; it is the explicit
operator-stability role of a per-cell somato-dendritic innovation measured from
paired neutral observations. The paper must nevertheless count and disclose
the neutral microphase. It cannot describe this variant as single-phase, and
must discuss proximity to contrastive/phase-separated local-learning methods.
## 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.
The formal justification is an application of the standard conditional-expectation projection
identity, not a new mathematical theorem. Among all predictors measurable from the chosen somatic
statistic, the neutral conditional mean uniquely minimizes residual nuisance power and leaves an
innovation orthogonal to every such predictor. The implemented per-cell affine model realizes the
restricted projection onto `{1,h_i}`. Positive norm matching cannot change the raw signal's cosine
direction at a fixed state, so the matched-raw ablation specifically distinguishes subtraction from
gain control. `THEORY.md` states the assumptions and the executable finite-sample identities.
## Claims currently allowed
- The inherited learned-feedback backbone preserves performance with depth better than fixed DFA
in the audited flattened-CIFAR MLP setting.
- Frozen controls show that per-cell residualization protects learning from soma-coupled traffic.
It mitigates but does not fully remove endogenous top-down traffic, so arbitrary contextual
innovation is not an allowed teaching-signal claim.
- Neutral residualization is the minimum-power nuisance removal within the predictor's somatic
function class. For the implemented diagonal affine class this is a narrow, testable statement,
not a claim of optimal population-level denoising.
- Simultaneous batched perturbations have a frozen hardware-independent advantage: K1/every4
retains 112.9% of the K16/every4 gain over DFA with 16x fewer logical loss queries, 11x less
calibration work, and 5.3x less total forward-equivalent work on CIFAR-10 d20/w64.
- Direct node perturbation solves the frozen useful-depth task while the amortized context
vectorizer does not. This identifies feedback amortization as the current engineering
bottleneck; it is not a novelty claim and the direct estimator's 68.4x work precludes presenting
it as the scalable algorithm.
- On the inherited reciprocal-KP substrate, dynamic paired-neutral innovation
first passes a frozen 20-epoch ResNet-20 validation gate at 83.58%, versus
82.66% for clean KP and 10% for the failed raw/norm-matched mixed-traffic
controls. The subsequent 200-epoch D3 validation reaches 91.18%, and the
independently frozen paired D4 test panel reaches 91.584% across seeds
10--14 versus clean KP's 91.388%, with a 0.131-point one-sided upper bound on
the clean-minus-dynamic deficit. This supports load-bearing innovation and
robustness on ResNet-20. It does not establish positive utility from adding
standard-network depth.
## 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.
- Dynamic neutral projection is a single-phase rule or has zero observation
cost merely because it adds no convolutional MAC.
- A single ResNet-20 architecture establishes scaling with added standard
depth, or licenses the sealed ResNet-32/56 panel after its oral-B
prerequisite failed.
- SDIL reproduces the complete Harnett population outcome-vectorization,
longitudinal, desired-velocity, or online-control signature. The untouched
oral-B R2 panel fails its joint gate despite successful task learning and
sign inversion.
- 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. The conditional-projection/SNR result and task-period absorption prediction are now proved and
executable. A broader paper would still benefit from a separately frozen failure test where
traffic depends on the teaching signal after conditioning on the available soma statistic.
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.
6. D3 and D4 now establish the narrow ResNet-20 result. A stronger scaling
contribution requires an independently frozen, genuinely depth-necessary
standard-network panel; the existing ResNet-20/32/56 panel remains sealed
because oral-B R2 failed.
7. A biological contribution beyond Harnett/BurstCCN requires a new mechanism
and prediction frozen independently of the failed R2 metrics. Successful
task learning, lesion sensitivity, and sign inversion alone cannot be
relabelled as the failed joint population-signature claim.
|
