early-epoch-clustering.md

Early-Epoch Clustering (SPARE)

Early-epoch clustering detects shortcut bias by exploiting the tendency of SGD-trained models to learn simple (often spurious) features early in training. By clustering representations or logits from the first few epochs, we can identify imbalanced clusters that suggest shortcut reliance.

When to Use

• You can capture early-epoch representations (logits or features).
• You lack protected-group labels but want a proxy signal.
• You want a training-time shortcut signal before full convergence.

Usage

from shortcut_detect.training import EarlyEpochClusteringDetector

# early_epoch_reps: array of shape (n_samples, n_features)
# labels optional; used only for cluster-label agreement

detector = EarlyEpochClusteringDetector(
    n_clusters=4,
    min_cluster_ratio=0.1,
    entropy_threshold=0.7,
)

detector.fit(early_epoch_reps, labels=labels, n_epochs=1)
print(detector.report_)

Interpretation

• Low entropy / low minority ratio → clusters are imbalanced, suggesting shortcut bias.
• High entropy / balanced clusters → no strong shortcut signal detected.

Reference

Yang et al. 2023, SPARE (SePArate early and REsample).