DTG-MA: Hard Guardrails Against Catastrophic Forgetting Through Attention Masking and Task Graphs

DTG-MA (Dynamic Task-Graph Masked Attention) is a novel approach to solving catastrophic forgetting in continual learning. It provides hard architectural guarantees against forgetting by using attention masking and task graphs to isolate task-specific computation pathways in neural networks.

Key features:
- Hard isolation: Uses masking in attention mechanisms to physically prevent interference between task-specific pathways
- Task graph structure: Each task gets dedicated edges in the computation graph, preventing parameter reuse and conflicts
- Memory efficient: Frozen edges for completed tasks, eliminating the need for replay buffers or storing old model copies
- GPU-friendly: Standard tensor operations map cleanly to accelerators without exotic requirements
- Interpretability: Easier to understand which task-specific subgraph activated for a given prediction

Compared to alternatives like EWC, LwF, and Replay, DTG-MA trades potential transfer learning for predictable, guaranteed isolation when task boundaries are known. Ideal for domain models, multi-tenant systems, and scenarios where data cannot be replayed.