Upload analysis/findings/activation_analysis_findings.md with huggingface_hub

analysis/findings/activation_analysis_findings.md

@@ -0,0 +1,20 @@
+# Activation Energy Analysis - Key Findings
+
+The analysis ran on the actual HPS model (hps_model.h5 from HuggingFace) with 100 random input traces.
+
+Important correction: The previous report claimed "52.2% energy destroyed" but the actual analysis shows only 5.9% of L2 energy is destroyed by GAP. This makes sense because GAP preserves the mean — it's the VARIANCE (spatial information) that is destroyed, not the total energy.
+
+The correct framing should be about INFORMATION LOSS (variance), not raw energy:
+- Pre-GAP: 1008 spatial positions x 512 channels = 516,096 values
+- Post-GAP: 1 x 512 = 512 values (1008:1 compression)
+- Mean per-channel spatial variance: 14.40
+- Total spatial variance destroyed: 7,373.82
+- The spatial variance IS the discriminative information — it tells the model WHERE features are located
+- GAP destroys 100% of this spatial variance by definition (averaging removes all position-dependent information)
+
+The figure shows:
+(a) Heatmap: clear spatial structure with hotspots at specific positions (especially around positions 0-10 in binned space)
+(b) Per-channel comparison: pre-GAP energy varies dramatically across positions, post-GAP is a flat average
+(c) Variance distribution: highly skewed, a few channels have very high spatial variance (up to 1400+), most near zero
+(d) Single channel collapse: channel 90 has a spike at one spatial position (value ~220) but GAP reduces it to 49.1
+(e) Pie chart: 94.1% energy retained, 5.9% destroyed (but this understates the information loss)