Limitations & Roadmap
We believe in transparency regarding the current state of the SCARCITY framework.
Current Limitations
1. Hardware Utilization
- GPU Acceleration: While the architecture supports GPU hooks, the current version primarily utilizes CPU for the core MPIE loops. GPU offloading is currently experimental.
- Memory Footprint: The hypergraph store can grow significantly with long-running sessions (500MB - 2GB), requiring periodic pruning.
2. Privacy Mechanisms
- Differential Privacy: The current implementation supports basic Gaussian noise addition ($\epsilon, \delta$). Advanced mechanisms like Secure Multi-Party Computation (SMPC) are not yet implemented.
Roadmap & Future Work
🚧 In Progress
- Advanced Privacy: Integrating homomorphic encryption for secure model aggregation.
- Distributed Simulation: Expanding the simulation engine to support distributed agent-based modeling across multiple nodes.
- Enhanced Meta-Learning: Improving the 5-tier meta-learning hierarchy for faster adaptation.
📋 Planned Features
- Full GPU Acceleration: moving tensor operations to CUDA for faster causal discovery.
- Kubernetes Support: Native Helm charts for deploying the Federation layer on K8s clusters.
- Model Export: Standardized ONNX export for causal graphs.