---

name: ewc_fisher_importance description: Estimate and normalize diagonal Fisher importance vectors for Elastic Weight Consolidation from first-order gradients.

EWC Fisher Importance

Use this skill after learning a previous task and before applying an EWC penalty. It converts per-example gradients into a nonnegative diagonal Fisher estimate and can compute normalized Fisher overlap diagnostics.

Inputs

• Per-example gradient vectors for the previous task.
• Optional Fisher vectors for normalization or overlap.

Outputs

• Mean squared-gradient Fisher vector.
• Trace-normalized Fisher vector.
• Fisher-overlap score for diagnostics.

Workflow

1. Collect gradients at or near the previous-task solution.
2. Square every gradient component.
3. Average squared gradients across examples.
4. Normalize by trace only for overlap/comparison diagnostics.
5. Pass the unnormalized Fisher to the EWC penalty.

Validation

Run:

python tests/test_fisher_importance.py

Limitations

The helper expects gradient vectors supplied by a model-specific training harness. It does not compute autograd gradients itself.