Spectral subspace initialization for nonlinear teachers / results.json
Mechanism confirmed, baseline not beaten
1{
2 "seed": 2170,
3 "population_identity": [
4 {
5 "rho": 0.25,
6 "alpha_mc": 0.5001306997076698,
7 "alpha_formula": 0.5001306997076698,
8 "ratio_to_rho": 2.000522798830679
9 },
10 {
11 "rho": 0.5,
12 "alpha_mc": 0.9946871386320814,
13 "alpha_formula": 0.9946871386320814,
14 "ratio_to_rho": 1.9893742772641627
15 },
16 {
17 "rho": 1.0,
18 "alpha_mc": 1.8529095525926127,
19 "alpha_formula": 1.8529095525926127,
20 "ratio_to_rho": 1.8529095525926127
21 },
22 {
23 "rho": 2.0,
24 "alpha_mc": 2.9509976010746466,
25 "alpha_formula": 2.9509976010746466,
26 "ratio_to_rho": 1.4754988005373233
27 }
28 ],
29 "null_population_alpha": 0.00024171174439707068,
30 "recovery_sweep": [
31 {
32 "rho": 0.5,
33 "n_over_d": 0.5,
34 "overlap_mean": 0.09060888690052565,
35 "overlap_std": 0.030134504526043913,
36 "top_gap": 1.5453272757704415
37 },
38 {
39 "rho": 0.5,
40 "n_over_d": 1,
41 "overlap_mean": 0.15656668553534497,
42 "overlap_std": 0.04817277473551983,
43 "top_gap": 1.4602187575170194
44 },
45 {
46 "rho": 0.5,
47 "n_over_d": 2,
48 "overlap_mean": 0.24480606936375426,
49 "overlap_std": 0.08194327430242403,
50 "top_gap": 0.6952277864751539
51 },
52 {
53 "rho": 0.5,
54 "n_over_d": 4,
55 "overlap_mean": 0.5115753809496275,
56 "overlap_std": 0.06774520721440508,
57 "top_gap": 0.5804667402414522
58 },
59 {
60 "rho": 0.5,
61 "n_over_d": 8,
62 "overlap_mean": 0.6917800427564864,
63 "overlap_std": 0.0467440118806628,
64 "top_gap": 0.4423015550081084
65 },
66 {
67 "rho": 1.0,
68 "n_over_d": 0.5,
69 "overlap_mean": 0.10277108597292317,
70 "overlap_std": 0.03526868842290766,
71 "top_gap": 4.759185208509549
72 },
73 {
74 "rho": 1.0,
75 "n_over_d": 1,
76 "overlap_mean": 0.10574578182918412,
77 "overlap_std": 0.04733190529544252,
78 "top_gap": 1.802418096001837
79 },
80 {
81 "rho": 1.0,
82 "n_over_d": 2,
83 "overlap_mean": 0.3107487411244205,
84 "overlap_std": 0.07968392110301836,
85 "top_gap": 1.0454289909812842
86 },
87 {
88 "rho": 1.0,
89 "n_over_d": 4,
90 "overlap_mean": 0.5153995718550086,
91 "overlap_std": 0.037731069902061196,
92 "top_gap": 0.7719655536378903
93 },
94 {
95 "rho": 1.0,
96 "n_over_d": 8,
97 "overlap_mean": 0.7244193981061101,
98 "overlap_std": 0.045572725153769526,
99 "top_gap": 1.1007279266867416
100 }
101 ],
102 "null_sweep": [
103 {
104 "n_over_d": 0.5,
105 "null_overlap_mean": 0.01371360209681287,
106 "null_overlap_std": 0.014927652275806978
107 },
108 {
109 "n_over_d": 1,
110 "null_overlap_mean": 0.004002377384768975,
111 "null_overlap_std": 0.0033523773930110028
112 },
113 {
114 "n_over_d": 2,
115 "null_overlap_mean": 0.011778719086513404,
116 "null_overlap_std": 0.012030336060758475
117 },
118 {
119 "n_over_d": 4,
120 "null_overlap_mean": 0.01719930717341046,
121 "null_overlap_std": 0.021867072517569024
122 },
123 {
124 "n_over_d": 8,
125 "null_overlap_mean": 0.02323170537203584,
126 "null_overlap_std": 0.025252580682295606
127 }
128 ],
129 "mlp_comparison": {
130 "available": true,
131 "device": "cuda",
132 "steps": 180,
133 "results": {
134 "xavier": {
135 "mean": 2.964919328689575,
136 "std": 0.3779699051041123
137 },
138 "spectral": {
139 "mean": 1.6825853983561199,
140 "std": 0.05065552326702389
141 }
142 }
143 }
144}