Spectral-gap adaptive polynomial filtering / bench_report.json
Failed on benchmark
1{
2 "bench_version": 1,
3 "track": "dynamics",
4 "model": "rnn_small",
5 "metric_direction": "lower is better",
6 "n_seeds": 8,
7 "baseline": {
8 "best_cfg": {
9 "lr": 0.01
10 },
11 "sweep": [
12 {
13 "cfg": {
14 "lr": 0.001
15 },
16 "mean": 0.004298779065720737
17 },
18 {
19 "cfg": {
20 "lr": 0.003
21 },
22 "mean": 0.0034615940821822733
23 },
24 {
25 "cfg": {
26 "lr": 0.01
27 },
28 "mean": 0.0014865802077110857
29 }
30 ],
31 "full": {
32 "mean": 0.0014198302378645167,
33 "std": 0.00039839114901047844,
34 "per_seed": [
35 0.0009327387670055032,
36 0.0018641706556081772,
37 0.0010785290505737066,
38 0.0020708823576569557,
39 0.0009260809747502208,
40 0.0014055331703275442,
41 0.0016443432541564107,
42 0.001436363672837615
43 ],
44 "n": 8
45 }
46 },
47 "idea": {
48 "mean": 0.0019242097914684564,
49 "std": 0.0006243480083057443,
50 "per_seed": [
51 0.0027711838483810425,
52 0.0015842666616663337,
53 0.0012298799119889736,
54 0.0026156860403716564,
55 0.001083912211470306,
56 0.002030868548899889,
57 0.002578162821009755,
58 0.0014997182879596949
59 ],
60 "n": 8
61 },
62 "comparison": {
63 "delta_mean": 0.0005043795536039397,
64 "idea_wins": 1,
65 "n_pairs": 8,
66 "per_seed_diffs": [
67 0.0018384450813755393,
68 -0.0002799039939418435,
69 0.000151350861415267,
70 0.0005448036827147007,
71 0.00015783123672008514,
72 0.0006253353785723448,
73 0.0009338195668533444,
74 6.335461512207985e-05
75 ],
76 "p_value": 0.05675,
77 "mde": 0.0005505516465549087,
78 "mde_rel_pct": 38.77587840240402,
79 "verdict": "no significant win",
80 "system_worked": false
81 },
82 "mechanism_signature": {
83 "prediction": "Jackson filtering suppresses the trained model output by p(2*gap-1) around its fixed point",
84 "signature": {
85 "gap_hat": 0.25,
86 "K": 7,
87 "sK": 1.75,
88 "predicted_fejer_gain": 0.173583984375,
89 "predicted_jackson_gain": 0.168701171875,
90 "observed_filtered_over_raw_norm": 0.7509765625,
91 "prediction_error_abs": 0.582275390625,
92 "confirmed": false
93 },
94 "idea_sweep": [
95 {
96 "cfg": {
97 "lr": 0.001
98 },
99 "result": {
100 "mean": 0.013434823020361364,
101 "std": 0.007597705589429824,
102 "per_seed": [
103 0.00517980195581913,
104 0.016346164047718048,
105 0.007690576370805502,
106 0.011888416483998299,
107 0.003950032871216536,
108 0.015627766028046608,
109 0.018083054572343826,
110 0.028712771832942963
111 ],
112 "n": 8
113 }
114 },
115 {
116 "cfg": {
117 "lr": 0.003
118 },
119 "result": {
120 "mean": 0.005034002184402198,
121 "std": 0.0020313352507723005,
122 "per_seed": [
123 0.004691103473305702,
124 0.00431825453415513,
125 0.003526391228660941,
126 0.01027995441108942,
127 0.004567801486700773,
128 0.0038863292429596186,
129 0.0040084864012897015,
130 0.0049936966970562935
131 ],
132 "n": 8
133 }
134 },
135 {
136 "cfg": {
137 "lr": 0.01
138 },
139 "result": {
140 "mean": 0.0019242097914684564,
141 "std": 0.0006243480083057443,
142 "per_seed": [
143 0.0027711838483810425,
144 0.0015842666616663337,
145 0.0012298799119889736,
146 0.0026156860403716564,
147 0.001083912211470306,
148 0.002030868548899889,
149 0.002578162821009755,
150 0.0014997182879596949
151 ],
152 "n": 8
153 }
154 }
155 ],
156 "math_checks": {
157 "3": {
158 "fejer_p1_error": 0.0,
159 "jackson_p1_error": 0.0,
160 "jackson_degree": 2,
161 "nonnegative": true
162 },
163 "7": {
164 "fejer_p1_error": 0.0,
165 "jackson_p1_error": 0.0,
166 "jackson_degree": 6,
167 "nonnegative": true
168 },
169 "15": {
170 "fejer_p1_error": 0.0,
171 "jackson_p1_error": 0.0,
172 "jackson_degree": 14,
173 "nonnegative": true
174 },
175 "iterate_identity_error": 3.7547848202023757e-17
176 },
177 "protocol": {
178 "epochs": 12,
179 "n_train": 400,
180 "n_test": 200,
181 "gap": 0.25,
182 "K": 7,
183 "matched_architecture": true,
184 "lr_union": [
185 {
186 "lr": 0.001
187 },
188 {
189 "lr": 0.003
190 },
191 {
192 "lr": 0.01
193 }
194 ]
195 }
196 }
197}