harihardiyan/graphene-ml: A Graphene-inspired physics-aware continual learning framework using hex-grid backbones and dynamic coercivity

Release: v1.0.0— Graphene-Inspired Continual Learning Engine This initial alpha release introduces Graphene ML, a specialized framework that leverages the topological and structural properties of graphene to solve the problem of catastrophic forgetting in sequential task learning. 🌟 Core Graphene Features Hexagonal Lattice Backbone: Implementation of a Hex-Grid Graph Convolutional Network (GCN) that treats feature relations as nodes in a honeycomb structure, providing a superior spatial prior for information propagation. Bio-Physical Edge Plasticity: A dynamic Hebbian-update mechanism for lattice "bonds" ($E$). The plasticity of these edges is locally modulated by Fisher Information, ensuring that critical structural connections are shielded from degradation. Dynamic Magnetic Coercivity (EWC-DC): A regularization module inspired by magnetic coercivity. It adaptively adjusts the "stiffness" of parameters ($C$) based on the local curvature of the loss surface, effectively locking in task-specific knowledge. Lattice Phase Transitions: Introduction of "Liquid" and "Solid" operational phases. The system maintains high entropy during learning (Liquid) and transitions to a frozen, stable state for inference (Solid). Structural Audit Suite: Integrated telemetry to monitor the lattice's physical health, including edge_energy (structural stability) and edge_drift (rate of topological change). 📊 Benchmark Performance (Split CIFAR-10) The framework has been validated using the standard 5-task sequential protocol: Average Final Accuracy: 83.7% (demonstrating significant resilience against forgetting). Knowledge Retention: Successfully maintained >90% accuracy on Task 1 even after completing the final task in the sequence. Backward Transfer: Observed stability in early task performance, indicating that the graphene lattice effectively accommodates new data without overwriting old representations. 🛠 Technical Specifications Encoder: ResNet-18 (Standard feature extractor). Backbone: 3-layer Graphene-Hex GCN. Optimization: Adam with Cosine Annealing schedule. Memory Management: Hybrid EWC-DC and a 5000-sample balanced experience replay. ⚠️ Current Limitations & Research Roadmap While this prototype demonstrates high stability, current research is focused on: Lattice Scaling: Optimizing the hex-grid density for higher resolution feature mapping. Hamiltonian Refining: Improving the precision of the underlying update rules to push average accuracy beyond the 90% threshold. Metaplasticity Tuning: Further refining the scaling factors ($\gamma$ and $\delta$) for diverse dataset textures. 🤝 Citation & Collaboration We encourage researchers to explore this lattice-based approach. If this implementation aids your work in Continual Learning or Physics-Informed Neural Networks, please cite the repository as outlined in the README.