The key idea: Instead of fixed weights, each weight exists in a "ring" of values that create a quantum-inspired superposition. Think of it as giving each weight multiple "perspectives" that combine during computation.

Results: - 10.6% accuracy with random weights (vs 8.8% for standard NN) - 90.1% accuracy after training (v2.0 released today) - Works specifically well on structured data like images

What makes it interesting: 1. No explicit convolution kernels, but achieves CNN-like performance 2. "Smart mirrors" implementation - O(1) memory per weight regardless of ring depth 3. Inspired by quantum superposition but runs on classical hardware

Technical details: - Each weight w becomes {w_center, w-δ, w+δ, ...} - Forward pass: y = 0.5×w_center + 0.25×w_left + 0.25×w_right - Training: gradients propagate through the ring structure

Code: https://github.com/Akbar1992A/ring-convolution-networks Paper: https://doi.org/10.5281/zenodo.15776775 PaperV2.0: https://doi.org/10.5281/zenodo.15777644

I'd love to hear your thoughts on: - Theoretical explanation for why this works on images - Potential applications beyond computer vision - Optimization techniques for ring structures

This is my first major research project, and I'm excited to share it with the HN community. Happy to answer any questions!