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
|
# Frozen Rain EP layer-state confirmation
This protocol was frozen after the single-seed S1 development screen in
`results/ep_bias/s1_summary.json`. S1 selected the only promoted corruption
ratio, `0.01`; stronger `0.1` and `4.0` conditions remain recorded failures.
## Fixed protocol
- Author implementation: `rain-neuromorphics/energy-based-learning`, revision
`6b253fd8a5d267535f58ab79992256ef10031ceb`.
- Model and learner: author `ConvHopfieldEnergy28` 32--64--10 with positive EP,
nudging `0.25`, 12 training relaxation iterations, 30 inference iterations,
author local parameter rules and SGD settings.
- Data: FashionMNIST training set only. Data seed `6100` fixes disjoint 10,000
training and 2,000 validation examples. The test set is not evaluated.
- Model/order seeds: `1989, 1990, 1991, 1992, 1993`; batch size 128; three
epochs; no epoch selection.
- Conditions: clean, raw structured bias, same-RMS zero-mean noise, constant
predictor, affine innovation predictor, and oracle subtraction.
- Bias: per-neuron affine function of the local first-phase state, normalized
by the experimenter to `0.01` times the initial clean layer-state-difference
RMS. This normalization is not visible to either predictor.
- Predictor: normalized local LMS at rate `0.2`. Constant and innovation see
the same 128 instruction-off observations in the existing first EP phase of
the first training minibatch. Both are then frozen. There is no extra
equilibrium phase and no backpropagation.
- Hardware: all six conditions for a seed run sequentially on one physical
GPU. Different seeds may run on GPUs 5 and 7 in parallel.
The injected bias is deliberately inside the affine predictor class. Passing
therefore confirms correction, causality and transfer to an EP implementation;
it is not independent evidence that a real device exposes the same feature.
## Frozen gate
Across all five paired seeds:
1. clean, noise, constant, innovation and oracle complete three finite epochs;
2. mean clean validation accuracy is at least 70%;
3. raw loses at least 10 accuracy points relative to clean in every seed;
4. innovation beats raw and constant in every seed, with 95% lower bounds of
at least 10 and 5 accuracy points respectively;
5. the 95% upper bounds on clean-minus-innovation, absolute
innovation-minus-oracle, and absolute noise-minus-clean are each below
three accuracy points;
6. innovation has lower final residual/clean state-difference RMS than constant
in every seed;
7. constant and innovation each use exactly 128 neutral observations;
8. innovation/clean mean wall-time ratio is at most `1.15`;
9. all records share one author revision, one SDIL revision, the fixed data
split and `autodiff_used_for_learning=false`.
Failure closes this exact confirmation. It does not authorize a new ratio,
predictor rate, seed replacement or extra calibration on the same holdout.
|
