Capacity-Shaped Binomial Bottleneck / results.json
Beats tuned baseline
1{
2 "device": "cuda",
3 "math_check": {
4 "max_abs_derivative_error": 1.013873429656087e-09,
5 "prior_sum": 0.9999999999999999,
6 "prior_endpoint_mass": 0.33637619018554665,
7 "prior": [
8 0.16818809509277333,
9 0.08809852600097658,
10 0.0695514678955081,
11 0.06136894226074231,
12 0.057277679443359375,
13 0.05551528930664039,
14 0.05551528930664034,
15 0.057277679443359375,
16 0.06136894226074226,
17 0.06955146789550798,
18 0.08809852600097658,
19 0.16818809509277333
20 ]
21 },
22 "baseline_continuous": {
23 "accuracy": 0.8755555748939514,
24 "sampled_count_accuracy": null,
25 "test_I_nats": 0.9117828607559204,
26 "test_KL_to_beta_binomial": 0.2944762110710144,
27 "test_endpoint_mass": 0.6830476522445679,
28 "train_ce_checkpoints": [
29 0.7499760571858931,
30 0.34990795623751686
31 ],
32 "activation_storage_bits_per_coordinate": 32.0
33 },
34 "capacity_shaped_binomial": {
35 "accuracy": 0.897777795791626,
36 "sampled_count_accuracy": 0.8955555558204651,
37 "test_I_nats": 0.9857069849967957,
38 "test_KL_to_beta_binomial": 0.09662292897701263,
39 "test_endpoint_mass": 0.5673930644989014,
40 "train_ce_checkpoints": [
41 0.8077772284934674,
42 0.2767513006394584
43 ],
44 "activation_storage_bits_per_coordinate": 3.169925001442312
45 }
46}