Modad Eval

modad-eval

Looking at Model Debiasing through the Lens of Anomaly Detection — Pastore et al. (2024) (arXiv:2407.17449, 2024)

What this evaluates

Evaluates a model's ability to mitigate spurious correlations (bias) in image classification by measuring performance on both overall test sets and specifically on bias-conflicting samples where the spurious attribute contradicts the true label.

Datasets

• Corrupted CIFAR-10 — total 60000; splits: train (-1), val (-1), test (-1)
• BAR — total 2595; splits: train (1941), test (654)
• BFFHQ — total 21200; splits: train (19200), val (1000), test (1000)
• Waterbirds — total ?; splits: train (-1), val (-1), test (-1)

Metrics

• Average Accuracy (primary) — range: percent
  • The fraction of correctly classified samples over the entire test set, which typically contains a balanced number of bias-aligned and conflicting samples.
• Conflicting Accuracy — range: percent
  • The fraction of correctly classified samples restricted strictly to the bias-conflicting subset of the test set.

Input / output format

Input: RGB image tensor paired with a ground-truth semantic class label.

Output: Predicted semantic class label.

Scoring recipe

def compute_metrics(predictions, gold_labels, is_conflicting_mask):
    avg_acc = (predictions == gold_labels).mean() * 100
    conflicting_acc = (predictions[is_conflicting_mask] == gold_labels[is_conflicting_mask]).mean() * 100
    return avg_acc, conflicting_acc

Common pitfalls

• BAR dataset lacks ground-truth bias annotations, so Conflicting Accuracy cannot be computed for it; only Average Accuracy is reported.
• Test sets for BFFHQ and Waterbirds are explicitly balanced (50% bias-aligned, 50% bias-conflicting), unlike the highly skewed training sets, which heavily influences Average Accuracy.
• Conflicting Accuracy must be computed strictly on samples where the spurious attribute contradicts the true semantic label; misidentifying bias-aligned samples as conflicting will artificially deflate this metric.

Evidence (verbatim from paper)

We employ two quantitative metrics in our results: Average Accuracy and Conflicting Accuracy. Average Accuracy is the accuracy over the whole test set, which typically contains a balanced number of bias-aligned and conflicting samples, except otherwise stated. Conflicting Accuracy refers to the average accuracy computed only on the bias-conflicting data.

Citation

@misc{pastore2024modad,
  title={Looking at Model Debiasing through the Lens of Anomaly Detection},
  author={Pastore et al. (2024)},
  year={2024},
  note={arXiv:2407.17449}
}

• arXiv: 2407.17449