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
|
{
"stage": "rain_ep_comparative32_beta_smoke_s4",
"status": "negative_beta_functional_under_author_protocol",
"protocol": {
"network": "author comparative-study ConvHopfieldEnergy32",
"dataset": "FashionMNIST",
"train_examples": 10000,
"evaluation_examples": 2000,
"epochs": 1,
"training_relaxation_iterations": 15,
"inference_relaxation_iterations": 60,
"schedule_epochs": 100,
"bias_ratio": 0.0
},
"rows": [
{
"beta_policy": "fixed_positive",
"training_accuracy": 0.2823,
"evaluation_accuracy": 0.614,
"wall_seconds": 129.7868938446045
},
{
"beta_policy": "fixed_negative",
"training_accuracy": 0.1553,
"evaluation_accuracy": 0.3155,
"wall_seconds": 130.82120490074158
},
{
"beta_policy": "random_sign",
"training_accuracy": 0.2316,
"evaluation_accuracy": 0.391,
"wall_seconds": 133.95944809913635,
"positive_steps": 36,
"negative_steps": 43
}
],
"interpretation": "Negative and random-sign beta learn above chance under the pinned author's comparative-study Conv32 architecture and hyperparameters. Their failure in the earlier Conv28 screen is protocol-specific. Random-sign beta must be included as a functioning same-cost baseline in the comparative32 bias matrix.",
"test_policy": "development test subset observed after one epoch; no accuracy claim"
}
|
