Identity-Paired Progressive Depth / results.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
 1{
 2  "seed": 1029,
 3  "math_predictions": [
 4    "C^{-1}(C(x))=x, so insertion loss change is zero up to floating point error.",
 5    "Composition roundoff should increase with coupling scale (products of f and g operators).",
 6    "After untie, a parameter step of size eta changes the pair output proportionally to eta (log-log slope 1)."
 7  ],
 8  "cancellation_sweep": [
 9    {
10      "scale": 0.01,
11      "max_abs_error": 2.384185791015625e-07,
12      "relative_error": 6.0507255780899e-09,
13      "predicted": "identity; fp error increases with operator products"
14    },
15    {
16      "scale": 0.03,
17      "max_abs_error": 2.384185791015625e-07,
18      "relative_error": 1.0371835612498303e-08,
19      "predicted": "identity; fp error increases with operator products"
20    },
21    {
22      "scale": 0.1,
23      "max_abs_error": 2.384185791015625e-07,
24      "relative_error": 1.6163652328020487e-08,
25      "predicted": "identity; fp error increases with operator products"
26    },
27    {
28      "scale": 0.3,
29      "max_abs_error": 2.384185791015625e-07,
30      "relative_error": 2.6485729196494805e-08,
31      "predicted": "identity; fp error increases with operator products"
32    },
33    {
34      "scale": 1.0,
35      "max_abs_error": 6.556510925292969e-07,
36      "relative_error": 6.519850746510443e-08,
37      "predicted": "identity; fp error increases with operator products"
38    }
39  ],
40  "untied_step_sweep": {
41    "gradient_norm": 0.07650767563339522,
42    "sweep": [
43      [
44        1e-05,
45        3.784688128217252e-07
46      ],
47      [
48        3e-05,
49        1.1330147494136314e-06
50      ],
51      [
52        0.0001,
53        3.7735932737830813e-06
54      ],
55      [
56        0.0003,
57        1.1323506324797927e-05
58      ],
59      [
60        0.001,
61        3.7741692688578436e-05
62      ]
63    ],
64    "loglog_slope": 0.999468080063024
65  },
66  "mini_experiment": {
67    "device": "cuda",
68    "before_insertion_loss": 0.001658762339502573,
69    "after_insertion_loss": 0.0016587624559178948,
70    "insertion_relative_change": 7.018203817060692e-08,
71    "paired_loss_after_300": 5.051885423662181e-14,
72    "shallow_loss_after_300": 1.8959316335311928e-14,
73    "random_deep_loss_after_300": 7.667316822335124e-05,
74    "paired_best_300": 2.020312492604856e-14,
75    "wall_seconds": 11.218648672103882
76  }
77}