Update README - GeoFractalDavid-Basin-k12 - Run 20251016_020120 - Acc 67.69%

README.md

@@ -14,7 +14,7 @@ metrics:
 - accuracy
 library_name: pytorch
 model-index:
-- name: GeoFractalDavid-Basin-k50
+- name: GeoFractalDavid-Basin-k12
   results:
   - task:
       type: image-classification
@@ -23,28 +23,27 @@ model-index:
       type: imagenet-1k
     metrics:
     - type: accuracy
-      value: 71.24
+      value: 67.69
       name: Validation Accuracy
 ---
 
-# GeoFractalDavid-Basin-k50: Geometric Basin Classification
+# GeoFractalDavid-Basin-k12: Geometric Basin Classification
 
 **GeoFractalDavid** achieves classification through geometric compatibility rather than cross-entropy.
 Features must "fit" geometric signatures: k-simplex shapes, Cantor positions, and hierarchical structure.
 
 ## 🎯 Performance
 
-- **Best Validation Accuracy**: 71.24%
-- **Epoch**: 10/10
-- **Training Time**: 22m 4s
+- **Best Validation Accuracy**: 67.69%
+- **Epoch**: 2/10
+- **Training Time**: 3m
 
 ### Per-Scale Performance
-- **Scale 448D**: 54.32%
-- **Scale 512D**: 52.39%
-- **Scale 576D**: 68.24%
-- **Scale 640D**: 50.34%
-- **Scale 704D**: 63.18%
-- **Scale 768D**: 46.11%
+- **Scale 384D**: 66.16%
+- **Scale 512D**: 66.40%
+- **Scale 768D**: 67.01%
+- **Scale 1024D**: 65.70%
+- **Scale 1280D**: 61.63%
 
 
 ## 🏗️ Architecture
@@ -54,9 +53,9 @@ Features must "fit" geometric signatures: k-simplex shapes, Cantor positions, an
 **Core Components**:
 - **Feature Dimension**: 512
 - **Number of Classes**: 1000
-- **k-Simplex Structure**: k=50 (51 vertices per class)
-- **Scales**: [448, 512, 576, 640, 704, 768]
-- **Total Simplex Vertices**: 51,000
+- **k-Simplex Structure**: k=12 (13 vertices per class)
+- **Scales**: [384, 512, 768, 1024, 1280]
+- **Total Simplex Vertices**: 13,000
 
 **Geometric Components**:
 1. **Feature Similarity**: Cosine similarity to k-simplex centroids
@@ -71,9 +70,9 @@ Each scale learns to weight these components differently.
 
 The alpha parameter controls middle-interval weighting in the Cantor staircase.
 
-- **Initial**: 0.3301
-- **Final**: -0.0637
-- **Change**: -0.3938
+- **Initial**: 0.3290
+- **Final**: 0.3158
+- **Change**: -0.0132
 - **Converged to 0.5**: False
 
 The Cantor staircase uses soft triadic decomposition with learnable alpha to map
@@ -84,35 +83,30 @@ features into [0,1] space with fractal structure.
 Each class has a learned scalar Cantor prototype. The model pulls features toward
 their class's Cantor position.
 
-**Scale 448D**:
-- Mean: 0.0248
-- Std: 0.0778
-- Range: [-0.1312, 0.1935]
+**Scale 384D**:
+- Mean: 0.2949
+- Std: 0.1159
+- Range: [0.0695, 0.4995]
 
 **Scale 512D**:
-- Mean: 0.0248
-- Std: 0.0778
-- Range: [-0.1313, 0.1933]
+- Mean: 0.2942
+- Std: 0.1160
+- Range: [0.0690, 0.4994]
 
-**Scale 576D**:
-- Mean: 0.0249
-- Std: 0.0778
-- Range: [-0.1309, 0.1934]
-
-**Scale 640D**:
-- Mean: 0.0247
-- Std: 0.0778
-- Range: [-0.1315, 0.1930]
+**Scale 768D**:
+- Mean: 0.3039
+- Std: 0.1147
+- Range: [0.0746, 0.5010]
 
-**Scale 704D**:
-- Mean: 0.0249
-- Std: 0.0778
-- Range: [-0.1309, 0.1935]
+**Scale 1024D**:
+- Mean: 0.2993
+- Std: 0.1153
+- Range: [0.0727, 0.4998]
 
-**Scale 768D**:
-- Mean: 0.0248
-- Std: 0.0779
-- Range: [-0.1314, 0.1931]
+**Scale 1280D**:
+- Mean: 0.2973
+- Std: 0.1156
+- Range: [0.0710, 0.4997]
 
 
 Most classes cluster around 0.5 (middle Cantor region), with smooth spread across [0,1].
@@ -122,12 +116,11 @@ This creates a continuous manifold rather than discrete bins.
 
 Each scale learns optimal weights for combining geometric components:
 
-**Scale 448D**: Feature=0.631, Cantor=0.022, Crystal=0.347
-**Scale 512D**: Feature=0.497, Cantor=0.023, Crystal=0.480
-**Scale 576D**: Feature=0.992, Cantor=0.001, Crystal=0.007
-**Scale 640D**: Feature=0.420, Cantor=0.023, Crystal=0.557
-**Scale 704D**: Feature=0.881, Cantor=0.001, Crystal=0.117
-**Scale 768D**: Feature=0.423, Cantor=0.013, Crystal=0.564
+**Scale 384D**: Feature=0.765, Cantor=0.070, Crystal=0.165
+**Scale 512D**: Feature=0.717, Cantor=0.072, Crystal=0.211
+**Scale 768D**: Feature=0.866, Cantor=0.030, Crystal=0.104
+**Scale 1024D**: Feature=0.744, Cantor=0.041, Crystal=0.215
+**Scale 1280D**: Feature=0.661, Cantor=0.042, Crystal=0.298
 
 
 **Pattern**: Lower scales rely on feature similarity, higher scales use crystal geometry.