| { |
| "alpha": 6, |
| "assessment": "verified", |
| "claim": 5, |
| "dimension": 256, |
| "direct_runs": [ |
| { |
| "alpha": 6, |
| "dimension": 256, |
| "linear_hidden_overlap": 0.026793740689754486, |
| "linear_test_loss_per_sample": 127.66737365722656, |
| "linear_visible_overlap": 0.8338882327079773, |
| "nonlinear_hidden_overlap": 0.06354930251836777, |
| "nonlinear_test_loss_per_sample": 128.17066955566406, |
| "nonlinear_visible_overlap": 0.6318047046661377, |
| "seed": 0, |
| "test_samples": 20000 |
| }, |
| { |
| "alpha": 6, |
| "dimension": 256, |
| "linear_hidden_overlap": 0.01279598567634821, |
| "linear_test_loss_per_sample": 127.60867309570312, |
| "linear_visible_overlap": 0.8281099796295166, |
| "nonlinear_hidden_overlap": 0.10018078237771988, |
| "nonlinear_test_loss_per_sample": 128.07933044433594, |
| "nonlinear_visible_overlap": 0.6960898041725159, |
| "seed": 1, |
| "test_samples": 20000 |
| }, |
| { |
| "alpha": 6, |
| "dimension": 256, |
| "linear_hidden_overlap": 0.039363835006952286, |
| "linear_test_loss_per_sample": 127.65738677978516, |
| "linear_visible_overlap": 0.8755549192428589, |
| "nonlinear_hidden_overlap": 0.07077246159315109, |
| "nonlinear_test_loss_per_sample": 128.1400604248047, |
| "nonlinear_visible_overlap": 0.7794100046157837, |
| "seed": 2, |
| "test_samples": 20000 |
| } |
| ], |
| "linear_loss_advantage_per_seed": [ |
| 0.5032958984375, |
| 0.4706573486328125, |
| 0.48267364501953125 |
| ], |
| "literal_claim": "Linear autoencoders achieve lower reconstruction (test) loss than nonlinear autoencoders yet fail to recover the hidden spike, demonstrating that self-supervised test loss misaligns with representation quality (Figure 2).", |
| "mean_linear_hidden_overlap": 0.026317853791018326, |
| "mean_linear_test_loss": 127.64447784423828, |
| "mean_nonlinear_hidden_overlap": 0.07816751549641292, |
| "mean_nonlinear_test_loss": 128.13002014160156, |
| "nonlinear_hidden_recovery_advantage_per_seed": [ |
| 0.03675556182861328, |
| 0.08738479670137167, |
| 0.03140862658619881 |
| ], |
| "pass": true, |
| "test_samples_per_seed": 20000 |
| } |
|
|