A-GEM gradient projection engine preventing negative interference and forgetting during continual task adaptation.
Learn more at GenPark and the GenPark MCP Catalog.
graph TD
A[Current Task Gradient g] --> B{g . g_ref >= 0?}
B -->|Yes| C[Apply Gradient Directly]
B -->|No: Interference| D[Orthogonal Projection g_proj = g - (g.g_ref / |g_ref|^2) * g_ref]
D --> E[g_proj . g_ref = 0: Zero Forgetting Guaranteed]
- Efficient orthogonal subspace projection.
- Eliminates backward transfer penalty.
- Pure Python standard library.