Update README - GeoFractalDavid-Basin-k4 - Run 20251015_230643 - Acc 82.36%

README.md

@@ -23,7 +23,7 @@ model-index:
       type: imagenet-1k
     metrics:
     - type: accuracy
-      value: 82.20
+      value: 82.36
       name: Validation Accuracy
 ---
 
@@ -34,18 +34,18 @@ Features must "fit" geometric signatures: k-simplex shapes, Cantor positions, an
 
 ## 🎯 Performance
 
-- **Best Validation Accuracy**: 82.20%
-- **Epoch**: 8/10
-- **Training Time**: 21m
+- **Best Validation Accuracy**: 82.36%
+- **Epoch**: 10/10
+- **Training Time**: 26m
 
 ### Per-Scale Performance
-- **Scale 576D**: 70.98%
-- **Scale 640D**: 71.06%
-- **Scale 704D**: 69.27%
-- **Scale 768D**: 81.15%
-- **Scale 832D**: 67.69%
-- **Scale 896D**: 70.00%
-- **Scale 960D**: 58.83%
+- **Scale 576D**: 69.88%
+- **Scale 640D**: 69.78%
+- **Scale 704D**: 67.68%
+- **Scale 768D**: 81.33%
+- **Scale 832D**: 65.97%
+- **Scale 896D**: 70.34%
+- **Scale 960D**: 59.12%
 
 
 ## 🏗️ Architecture
@@ -73,8 +73,8 @@ Each scale learns to weight these components differently.
 The alpha parameter controls middle-interval weighting in the Cantor staircase.
 
 - **Initial**: 0.4817
-- **Final**: 0.0420
-- **Change**: -0.4397
+- **Final**: 0.0247
+- **Change**: -0.4570
 - **Converged to 0.5**: False
 
 The Cantor staircase uses soft triadic decomposition with learnable alpha to map
@@ -86,39 +86,39 @@ Each class has a learned scalar Cantor prototype. The model pulls features towar
 their class's Cantor position.
 
 **Scale 576D**:
-- Mean: 0.0870
-- Std: 0.1245
-- Range: [-0.0823, 0.4752]
+- Mean: 0.0753
+- Std: 0.1164
+- Range: [-0.0998, 0.4649]
 
 **Scale 640D**:
-- Mean: 0.0869
-- Std: 0.1244
-- Range: [-0.0818, 0.4751]
+- Mean: 0.0751
+- Std: 0.1163
+- Range: [-0.0999, 0.4648]
 
 **Scale 704D**:
-- Mean: 0.0869
-- Std: 0.1244
-- Range: [-0.0821, 0.4751]
+- Mean: 0.0753
+- Std: 0.1164
+- Range: [-0.0999, 0.4648]
 
 **Scale 768D**:
-- Mean: 0.0911
-- Std: 0.1275
-- Range: [-0.0765, 0.4779]
+- Mean: 0.0787
+- Std: 0.1191
+- Range: [-0.0950, 0.4677]
 
 **Scale 832D**:
-- Mean: 0.0867
-- Std: 0.1243
-- Range: [-0.0820, 0.4750]
+- Mean: 0.0750
+- Std: 0.1163
+- Range: [-0.0999, 0.4647]
 
 **Scale 896D**:
-- Mean: 0.0891
-- Std: 0.1260
-- Range: [-0.0787, 0.4766]
+- Mean: 0.0771
+- Std: 0.1179
+- Range: [-0.0978, 0.4663]
 
 **Scale 960D**:
-- Mean: 0.0880
-- Std: 0.1251
-- Range: [-0.0800, 0.4757]
+- Mean: 0.0761
+- Std: 0.1170
+- Range: [-0.0991, 0.4654]
 
 
 Most classes cluster around 0.5 (middle Cantor region), with smooth spread across [0,1].
@@ -128,13 +128,13 @@ This creates a continuous manifold rather than discrete bins.
 
 Each scale learns optimal weights for combining geometric components:
 
-**Scale 576D**: Feature=0.780, Cantor=0.030, Crystal=0.189
-**Scale 640D**: Feature=0.734, Cantor=0.032, Crystal=0.234
-**Scale 704D**: Feature=0.642, Cantor=0.034, Crystal=0.325
-**Scale 768D**: Feature=0.989, Cantor=0.001, Crystal=0.009
-**Scale 832D**: Feature=0.619, Cantor=0.028, Crystal=0.353
-**Scale 896D**: Feature=0.769, Cantor=0.003, Crystal=0.229
-**Scale 960D**: Feature=0.300, Cantor=0.005, Crystal=0.695
+**Scale 576D**: Feature=0.791, Cantor=0.028, Crystal=0.182
+**Scale 640D**: Feature=0.741, Cantor=0.029, Crystal=0.230
+**Scale 704D**: Feature=0.640, Cantor=0.031, Crystal=0.329
+**Scale 768D**: Feature=0.991, Cantor=0.001, Crystal=0.008
+**Scale 832D**: Feature=0.613, Cantor=0.026, Crystal=0.362
+**Scale 896D**: Feature=0.772, Cantor=0.002, Crystal=0.226
+**Scale 960D**: Feature=0.252, Cantor=0.004, Crystal=0.743
 
 
 **Pattern**: Lower scales rely on feature similarity, higher scales use crystal geometry.