Create 1_1_2d_scattering_kyumatio_decomposition_analysis.txt

spectral/experiment_1/1_1_2d_scattering_kyumatio_decomposition_analysis.txt

@@ -0,0 +1,277 @@
+K=243, spatial=8x8, flat=15552
+
+Model params: 156,016
+
+======================================================================
+SCATTERING GEOMETRIC AUTOPSY — 90 epochs
+======================================================================
+
+[Pre-training] Collecting features...
+[Pre-training] Measuring augmentation stability...
+  Aug stability: cos=0.5737±0.2524
+
+  E  1: train=47.0% val=57.8% loss=2.0307 (6s) *
+    CV: raw=0.2945 bn=0.2486 | o0=0.3275 o1=0.2489 o2=0.3122
+    SVD eff_dim: raw=128.8 bn=1031.9 | dim90: raw=31 bn=545
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3352 o1=0.0024 o2=0.0124
+    Class sep (bn): cos=0.6583±0.2076
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+  E  2: train=52.9% val=60.5% loss=1.7034 (6s) *
+  E  3: train=55.2% val=61.0% loss=1.4258 (6s) *
+  E  4: train=57.2% val=64.1% loss=1.2565 (6s) *
+
+  E  5: train=58.5% val=64.8% loss=1.2155 (6s) *
+    CV: raw=0.2749 bn=0.2244 | o0=0.3554 o1=0.2772 o2=0.3297
+    SVD eff_dim: raw=128.8 bn=917.9 | dim90: raw=31 bn=441
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3317 o1=0.0025 o2=0.0123
+    Class sep (bn): cos=0.6375±0.2007
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+  E  6: train=58.6% val=64.2% loss=1.2109 (7s)
+  E  7: train=58.8% val=64.5% loss=1.2015 (6s)
+  E  8: train=59.4% val=62.7% loss=1.1907 (6s)
+  E  9: train=59.3% val=63.6% loss=1.1898 (6s)
+
+  E 10: train=59.5% val=65.6% loss=1.1816 (6s) *
+    CV: raw=0.3180 bn=0.2146 | o0=0.3435 o1=0.2676 o2=0.3504
+    SVD eff_dim: raw=128.8 bn=914.8 | dim90: raw=31 bn=439
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3289 o1=0.0025 o2=0.0122
+    Class sep (bn): cos=0.6110±0.2160
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+  E 11: train=59.6% val=64.5% loss=1.1859 (6s)
+  E 12: train=59.4% val=63.9% loss=1.1870 (6s)
+  E 13: train=59.4% val=63.9% loss=1.1857 (6s)
+  E 14: train=59.4% val=63.5% loss=1.1853 (6s)
+
+  E 15: train=59.8% val=65.4% loss=1.1814 (6s)
+    CV: raw=0.2924 bn=0.2356 | o0=0.3495 o1=0.2712 o2=0.3477
+    SVD eff_dim: raw=128.8 bn=925.4 | dim90: raw=31 bn=449
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3350 o1=0.0025 o2=0.0122
+    Class sep (bn): cos=0.6217±0.2141
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+  E 16: train=59.8% val=65.0% loss=1.1753 (6s)
+  E 17: train=59.7% val=64.7% loss=1.1749 (6s)
+  E 18: train=59.8% val=65.0% loss=1.1794 (6s)
+  E 19: train=59.5% val=64.8% loss=1.1829 (6s)
+
+  E 20: train=59.5% val=63.6% loss=1.1803 (6s)
+    CV: raw=0.2897 bn=0.2294 | o0=0.3491 o1=0.2905 o2=0.3165
+    SVD eff_dim: raw=128.8 bn=909.3 | dim90: raw=31 bn=436
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3403 o1=0.0025 o2=0.0125
+    Class sep (bn): cos=0.6000±0.2192
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+  E 21: train=63.6% val=68.2% loss=1.0752 (6s) *
+  E 22: train=64.4% val=68.2% loss=1.0519 (6s) *
+  E 23: train=64.5% val=68.2% loss=1.0494 (6s)
+  E 24: train=64.8% val=68.2% loss=1.0414 (6s)
+
+  E 25: train=65.3% val=68.2% loss=1.0326 (6s)
+    CV: raw=0.2907 bn=0.2560 | o0=0.3598 o1=0.2710 o2=0.3416
+    SVD eff_dim: raw=128.8 bn=932.9 | dim90: raw=31 bn=453
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3332 o1=0.0024 o2=0.0124
+    Class sep (bn): cos=0.6274±0.2120
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+  E 26: train=64.7% val=68.1% loss=1.0355 (6s)
+  E 27: train=65.1% val=68.5% loss=1.0296 (6s) *
+  E 28: train=65.1% val=68.8% loss=1.0271 (6s) *
+  E 29: train=65.3% val=68.2% loss=1.0277 (6s)
+
+  E 30: train=65.4% val=68.5% loss=1.0227 (6s)
+    CV: raw=0.3282 bn=0.2305 | o0=0.3791 o1=0.3095 o2=0.3210
+    SVD eff_dim: raw=128.8 bn=936.2 | dim90: raw=31 bn=456
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3385 o1=0.0025 o2=0.0126
+    Class sep (bn): cos=0.6251±0.2145
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+  E 31: train=65.3% val=69.4% loss=1.0239 (6s) *
+  E 32: train=65.3% val=69.6% loss=1.0216 (6s) *
+  E 33: train=65.2% val=68.7% loss=1.0230 (6s)
+  E 34: train=65.5% val=69.1% loss=1.0179 (6s)
+
+  E 35: train=65.6% val=69.4% loss=1.0156 (7s)
+    CV: raw=0.3007 bn=0.2125 | o0=0.3542 o1=0.3097 o2=0.3246
+    SVD eff_dim: raw=128.8 bn=940.8 | dim90: raw=31 bn=460
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3401 o1=0.0024 o2=0.0124
+    Class sep (bn): cos=0.6180±0.2179
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+  E 36: train=65.4% val=69.6% loss=1.0177 (6s)
+  E 37: train=65.7% val=69.5% loss=1.0125 (6s)
+  E 38: train=65.5% val=69.2% loss=1.0168 (6s)
+  E 39: train=65.8% val=69.0% loss=1.0168 (6s)
+
+  E 40: train=65.7% val=69.1% loss=1.0156 (6s)
+    CV: raw=0.3044 bn=0.2387 | o0=0.3140 o1=0.2671 o2=0.3500
+    SVD eff_dim: raw=128.8 bn=942.4 | dim90: raw=31 bn=461
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3368 o1=0.0025 o2=0.0126
+    Class sep (bn): cos=0.6256±0.2150
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+  E 41: train=66.8% val=70.0% loss=0.9855 (6s) *
+  E 42: train=67.0% val=70.1% loss=0.9793 (6s) *
+  E 43: train=67.1% val=70.1% loss=0.9800 (6s)
+  E 44: train=67.0% val=70.0% loss=0.9805 (6s)
+
+  E 45: train=67.1% val=70.0% loss=0.9790 (6s)
+    CV: raw=0.2830 bn=0.2437 | o0=0.3386 o1=0.3004 o2=0.3423
+    SVD eff_dim: raw=128.8 bn=943.9 | dim90: raw=31 bn=462
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3370 o1=0.0025 o2=0.0126
+    Class sep (bn): cos=0.6170±0.2182
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+  E 46: train=67.2% val=70.1% loss=0.9746 (6s) *
+  E 47: train=67.4% val=70.5% loss=0.9704 (6s) *
+  E 48: train=67.4% val=70.0% loss=0.9761 (6s)
+  E 49: train=67.4% val=70.1% loss=0.9715 (6s)
+
+  E 50: train=67.0% val=70.6% loss=0.9774 (6s) *
+    CV: raw=0.3256 bn=0.2417 | o0=0.3418 o1=0.2891 o2=0.3307
+    SVD eff_dim: raw=128.8 bn=942.7 | dim90: raw=31 bn=460
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3290 o1=0.0024 o2=0.0123
+    Class sep (bn): cos=0.6146±0.2195
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+  E 51: train=67.4% val=70.5% loss=0.9721 (6s)
+  E 52: train=67.5% val=70.3% loss=0.9668 (6s)
+  E 53: train=67.3% val=70.6% loss=0.9740 (6s)
+  E 54: train=67.4% val=70.4% loss=0.9736 (6s)
+
+  E 55: train=67.2% val=70.3% loss=0.9716 (6s)
+    CV: raw=0.2874 bn=0.2374 | o0=0.3227 o1=0.3073 o2=0.3206
+    SVD eff_dim: raw=128.8 bn=945.3 | dim90: raw=31 bn=463
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3392 o1=0.0024 o2=0.0123
+    Class sep (bn): cos=0.6195±0.2179
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+  E 56: train=67.4% val=70.5% loss=0.9723 (6s)
+  E 57: train=67.2% val=70.2% loss=0.9717 (6s)
+  E 58: train=67.2% val=69.8% loss=0.9714 (6s)
+  E 59: train=67.2% val=70.3% loss=0.9699 (6s)
+
+  E 60: train=67.6% val=70.5% loss=0.9684 (6s)
+    CV: raw=0.3094 bn=0.2229 | o0=0.3546 o1=0.2708 o2=0.3450
+    SVD eff_dim: raw=128.8 bn=945.7 | dim90: raw=31 bn=463
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3330 o1=0.0025 o2=0.0124
+    Class sep (bn): cos=0.6100±0.2213
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+  E 61: train=67.7% val=70.5% loss=0.9599 (6s)
+  E 62: train=67.7% val=70.3% loss=0.9627 (6s)
+  E 63: train=67.7% val=70.6% loss=0.9594 (6s)
+  E 64: train=67.8% val=70.3% loss=0.9598 (6s)
+
+  E 65: train=68.0% val=70.8% loss=0.9582 (6s) *
+    CV: raw=0.2935 bn=0.2245 | o0=0.3507 o1=0.2660 o2=0.3206
+    SVD eff_dim: raw=128.8 bn=945.0 | dim90: raw=31 bn=462
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3290 o1=0.0024 o2=0.0121
+    Class sep (bn): cos=0.6082±0.2214
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+  E 66: train=67.9% val=70.5% loss=0.9602 (6s)
+  E 67: train=67.6% val=70.7% loss=0.9591 (6s)
+  E 68: train=67.7% val=70.6% loss=0.9626 (6s)
+  E 69: train=68.0% val=70.5% loss=0.9562 (6s)
+
+  E 70: train=67.8% val=70.5% loss=0.9584 (6s)
+    CV: raw=0.2901 bn=0.2361 | o0=0.3693 o1=0.2753 o2=0.3459
+    SVD eff_dim: raw=128.8 bn=945.9 | dim90: raw=31 bn=463
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3395 o1=0.0025 o2=0.0124
+    Class sep (bn): cos=0.6277±0.2143
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+  E 71: train=67.7% val=70.6% loss=0.9592 (6s)
+  E 72: train=67.8% val=70.7% loss=0.9608 (6s)
+  E 73: train=67.6% val=70.8% loss=0.9637 (6s) *
+  E 74: train=67.7% val=70.7% loss=0.9589 (6s)
+
+  E 75: train=67.5% val=70.6% loss=0.9617 (6s)
+    CV: raw=0.2650 bn=0.2113 | o0=0.3409 o1=0.2532 o2=0.3204
+    SVD eff_dim: raw=128.8 bn=945.1 | dim90: raw=31 bn=462
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3342 o1=0.0025 o2=0.0125
+    Class sep (bn): cos=0.6311±0.2128
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+  E 76: train=67.7% val=70.7% loss=0.9598 (6s)
+  E 77: train=67.5% val=70.7% loss=0.9618 (6s)
+  E 78: train=68.0% val=70.6% loss=0.9581 (6s)
+  E 79: train=67.7% val=70.6% loss=0.9565 (6s)
+
+  E 80: train=67.8% val=70.5% loss=0.9606 (6s)
+    CV: raw=0.3145 bn=0.2327 | o0=0.3830 o1=0.3325 o2=0.3248
+    SVD eff_dim: raw=128.8 bn=945.8 | dim90: raw=31 bn=463
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3355 o1=0.0025 o2=0.0124
+    Class sep (bn): cos=0.6369±0.2100
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+  E 81: train=67.8% val=70.6% loss=0.9588 (6s)
+  E 82: train=67.8% val=70.5% loss=0.9551 (6s)
+  E 83: train=67.9% val=70.6% loss=0.9587 (6s)
+  E 84: train=67.8% val=70.4% loss=0.9594 (6s)
+
+  E 85: train=67.8% val=70.4% loss=0.9570 (6s)
+    CV: raw=0.2825 bn=0.2443 | o0=0.3515 o1=0.2843 o2=0.3332
+    SVD eff_dim: raw=128.8 bn=945.3 | dim90: raw=31 bn=462
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3361 o1=0.0025 o2=0.0125
+    Class sep (bn): cos=0.6358±0.2108
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+  E 86: train=67.9% val=70.5% loss=0.9585 (6s)
+  E 87: train=67.8% val=70.6% loss=0.9553 (6s)
+  E 88: train=67.8% val=70.6% loss=0.9608 (6s)
+  E 89: train=68.0% val=70.7% loss=0.9579 (6s)
+
+  E 90: train=67.8% val=70.5% loss=0.9584 (6s)
+    CV: raw=0.2696 bn=0.2400 | o0=0.3435 o1=0.2720 o2=0.2977
+    SVD eff_dim: raw=128.8 bn=946.3 | dim90: raw=31 bn=463
+    Raw L2: o0=7.148 o1=3.108 o2=11.854
+    BN var: o0=0.3389 o1=0.0026 o2=0.0124
+    Class sep (bn): cos=0.6420±0.2075
+    Procrustes o0↔o1: align=0.1511 o1↔o2: align=0.3430
+
+======================================================================
+FINAL RESULTS
+======================================================================
+  Best val accuracy: 70.8%
+  Augmentation stability: cos=0.5737±0.2524
+
+  Full diagnostics saved to analysis/scattering_autopsy.json
+
+======================================================================
+GEOMETRIC THEORY EXTRACTION
+======================================================================
+
+  SCATTERING FEATURE GEOMETRY:
+    Order 0 (3 ch):   L2=7.148, BN_var=0.3389
+    Order 1 (48 ch):  L2=3.108, BN_var=0.0026  
+    Order 2 (192 ch): L2=11.854, BN_var=0.0124
+    Scale ratio o0/o2: 0.6x
+    
+  EFFECTIVE DIMENSIONALITY:
+    Raw features: eff_dim=128.8, dim_90=31, dim_99=217
+    After BN:     eff_dim=946.3, dim_90=463, dim_99=2276
+    
+  HYPERSPHERE GEOMETRY (CV):
+    Raw:      0.2696
+    After BN: 0.2400
+    Target:   0.20 (GeoLIP universal attractor)
+    
+  CLASS SEPARATION:
+    Raw:  cos=0.6611±0.2304
+    BN:   cos=0.6420±0.2075
+    
+  INTER-ORDER PROCRUSTES:
+    o0↔o1: alignment=0.1511
+    o1↔o2: alignment=0.3430
+    
+  AUGMENTATION STABILITY:
+    cos=0.5737±0.2524
+    (1.0=perfectly stable, 0.0=completely random)
+
+======================================================================
+DONE — Full record in analysis/scattering_autopsy.json
+======================================================================