IncDet‑EWC (Huber EWC + Pseudo‑Annotation)

[maniq227]

Title: IncDet‑EWC: add Huber EWC + Pseudo‑Annotation, tests, examples, and how‑to docs

Summary

• Introduces HuberEWCPlugin and PseudoAnnotationPlugin for classification CL.
• Adds unit + integration tests, runnable examples, and a GitBook how‑to guide.

Changes

• Code: avalanche/training/plugins/incdet_ewc.py, avalanche/training/plugins/pseudo_annotation.py
• Exports: avalanche/training/plugins/__init__.py
• Tests: tests/plugins/test_incdet_ewc_plugin.py, tests/plugins/test_pseudo_annotation_plugin.py, tests/strategies/test_incdet_ewc_integration.py
• Examples: examples/incdet_ewc_cifar.py, examples/incdet_ewc_splitmnist.py, run_incdet_ewc.py
• Docs (GitBook): docs/gitbook/how-tos/incdet_ewc.md + link in docs/gitbook/SUMMARY.md

Motivation

• Huber variant of EWC improves stability at larger regularization strengths.
• Minimal pseudo‑annotation augments current experience with confident predictions on previously seen classes.
• Adds a short, task‑oriented guide to complement API docs.

How to run

# SplitMNIST quickstart
python run_incdet_ewc.py --device auto --train_mb_size 64 --eval_mb_size 64 \
  --train_epochs 1 --lr 0.01 --momentum 0.9 --ewc_lambda 1000.0 --beta 0.5 \
  --confidence_thresh 0.95 --n_experiences 2 --seed 42

# CIFAR
python examples/incdet_ewc_cifar.py --benchmark cifar10 --use_huber 1 --train_mb_size 128 --train_epochs 2

Notes

• Sphinx API doc already lists the plugins; this PR adds a GitBook how‑to (quickstart + minimal API snippet).
• No breaking changes; plugins are opt‑in via plugins=[...].

Checklist

• CI green
• Docs build
• Examples run
• Review feedback addressed