🚀 CascadeScaleNetSurge

Variance-driven EMA controller for MultiHead CNNs with adaptive thresholds and dynamic load balancing via freeze, scale and sigmoid regulation.

Author: Bato Naidanov (2025)

This repository demonstrates how to dynamically regulate the computational load of MultiHead CNNs based on real-time variance tracking, using a universal controller capable of freeze masks, smooth scaling and sigmoid load tuning.

📌 Key Features

• MultiHead CNN architecture with independent heads.
• EMA variance tracking per head with adaptive thresholds.
• Universal controller supporting:
  • freeze (requires_grad=False)
  • scale (soft scaling of activations)
  • sigmoid_scale (smooth scale ~0.5-1.0)
  • hard_mask (binary 0/1)
• Heatmaps and scatter plots for head activity analysis.
• CSV summaries of accuracy, variance, scale, time.

🚀 How to run

pip install -r requirements.txt
python examples/quick_test.py

🔖 License

Licensed under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). (C) Bato Naidanov, 2025