Update README - GeoFractalDavid-Basin-k50 - Run 20251015_222757 - Acc 79.01%

README.md

@@ -14,7 +14,7 @@ metrics:
 - accuracy
 library_name: pytorch
 model-index:
-- name: GeoFractalDavid-Basin-k100
+- name: GeoFractalDavid-Basin-k50
   results:
   - task:
       type: image-classification
@@ -23,27 +23,29 @@ model-index:
       type: imagenet-1k
     metrics:
     - type: accuracy
-      value: 82.12
+      value: 79.01
       name: Validation Accuracy
 ---
 
-# GeoFractalDavid-Basin-k100: Geometric Basin Classification
+# GeoFractalDavid-Basin-k50: 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**: 82.12%
-- **Epoch**: 10/10
-- **Training Time**: 31m 21s
+- **Best Validation Accuracy**: 79.01%
+- **Epoch**: 2/10
+- **Training Time**: 5m
 
 ### Per-Scale Performance
-- **Scale 384D**: 70.35%
-- **Scale 512D**: 67.95%
-- **Scale 768D**: 80.83%
-- **Scale 1024D**: 68.19%
-- **Scale 1280D**: 69.43%
+- **Scale 576D**: 78.05%
+- **Scale 640D**: 77.73%
+- **Scale 704D**: 78.18%
+- **Scale 768D**: 77.52%
+- **Scale 832D**: 77.78%
+- **Scale 896D**: 77.79%
+- **Scale 960D**: 77.53%
 
 
 ## 🏗️ Architecture
@@ -53,9 +55,9 @@ Features must "fit" geometric signatures: k-simplex shapes, Cantor positions, an
 **Core Components**:
 - **Feature Dimension**: 768
 - **Number of Classes**: 1000
-- **k-Simplex Structure**: k=100 (101 vertices per class)
-- **Scales**: [384, 512, 768, 1024, 1280]
-- **Total Simplex Vertices**: 101,000
+- **k-Simplex Structure**: k=50 (51 vertices per class)
+- **Scales**: [576, 640, 704, 768, 832, 896, 960]
+- **Total Simplex Vertices**: 51,000
 
 **Geometric Components**:
 1. **Feature Similarity**: Cosine similarity to k-simplex centroids
@@ -70,9 +72,9 @@ Each scale learns to weight these components differently.
 
 The alpha parameter controls middle-interval weighting in the Cantor staircase.
 
-- **Initial**: 0.4765
-- **Final**: -0.0067
-- **Change**: -0.4833
+- **Initial**: 0.4746
+- **Final**: 0.3957
+- **Change**: -0.0789
 - **Converged to 0.5**: False
 
 The Cantor staircase uses soft triadic decomposition with learnable alpha to map
@@ -83,30 +85,40 @@ 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 384D**:
-- Mean: 0.0564
-- Std: 0.1033
-- Range: [-0.1000, 0.4476]
+**Scale 576D**:
+- Mean: 0.4079
+- Std: 0.1335
+- Range: [0.0001, 0.5663]
 
-**Scale 512D**:
-- Mean: 0.0561
-- Std: 0.1032
-- Range: [-0.1000, 0.4475]
+**Scale 640D**:
+- Mean: 0.4078
+- Std: 0.1341
+- Range: [-0.0002, 0.5663]
+
+**Scale 704D**:
+- Mean: 0.4076
+- Std: 0.1266
+- Range: [0.0002, 0.5618]
 
 **Scale 768D**:
-- Mean: 0.0583
-- Std: 0.1049
-- Range: [-0.1000, 0.4491]
+- Mean: 0.4047
+- Std: 0.1384
+- Range: [-0.0000, 0.5645]
+
+**Scale 832D**:
+- Mean: 0.4062
+- Std: 0.1359
+- Range: [-0.0002, 0.5651]
 
-**Scale 1024D**:
-- Mean: 0.0571
-- Std: 0.1039
-- Range: [-0.1000, 0.4482]
+**Scale 896D**:
+- Mean: 0.4052
+- Std: 0.1302
+- Range: [-0.0001, 0.5604]
 
-**Scale 1280D**:
-- Mean: 0.0571
-- Std: 0.1040
-- Range: [-0.1000, 0.4482]
+**Scale 960D**:
+- Mean: 0.4056
+- Std: 0.1329
+- Range: [-0.0002, 0.5615]
 
 
 Most classes cluster around 0.5 (middle Cantor region), with smooth spread across [0,1].
@@ -116,11 +128,13 @@ This creates a continuous manifold rather than discrete bins.
 
 Each scale learns optimal weights for combining geometric components:
 
-**Scale 384D**: Feature=0.830, Cantor=0.021, Crystal=0.148
-**Scale 512D**: Feature=0.704, Cantor=0.022, Crystal=0.274
-**Scale 768D**: Feature=0.977, Cantor=0.001, Crystal=0.022
-**Scale 1024D**: Feature=0.877, Cantor=0.003, Crystal=0.119
-**Scale 1280D**: Feature=0.677, Cantor=0.003, Crystal=0.320
+**Scale 576D**: Feature=0.642, Cantor=0.058, Crystal=0.301
+**Scale 640D**: Feature=0.621, Cantor=0.058, Crystal=0.321
+**Scale 704D**: Feature=0.836, Cantor=0.027, Crystal=0.137
+**Scale 768D**: Feature=0.545, Cantor=0.071, Crystal=0.383
+**Scale 832D**: Feature=0.606, Cantor=0.057, Crystal=0.337
+**Scale 896D**: Feature=0.808, Cantor=0.028, Crystal=0.163
+**Scale 960D**: Feature=0.713, Cantor=0.039, Crystal=0.248
 
 
 **Pattern**: Lower scales rely on feature similarity, higher scales use crystal geometry.