diff options
| author | YurenHao0426 <Blackhao0426@gmail.com> | 2026-08-06 17:30:19 -0500 |
|---|---|---|
| committer | YurenHao0426 <Blackhao0426@gmail.com> | 2026-08-06 17:30:19 -0500 |
| commit | 53e83e9e68618174b6c9f5e4927463c5ee45434e (patch) | |
| tree | f43b775097d6f5948fddd1a7d7b6ad8bdab728b1 /results/ep_bias/s4_summary.json | |
| parent | c868f021850d4f8a91a4eca97ad475aa6fafdc7d (diff) | |
results: validate Rain comparative beta policies
Diffstat (limited to 'results/ep_bias/s4_summary.json')
| -rw-r--r-- | results/ep_bias/s4_summary.json | 39 |
1 files changed, 39 insertions, 0 deletions
diff --git a/results/ep_bias/s4_summary.json b/results/ep_bias/s4_summary.json new file mode 100644 index 0000000..32db33d --- /dev/null +++ b/results/ep_bias/s4_summary.json @@ -0,0 +1,39 @@ +{ + "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" +} |
