darcook/DCPV-Artifacts

DCPV: Density-Clustered / OOD-Aware Predictive Variance Artifacts

This repository contains the experimental artifacts, results, and datasets for the paper "Mitigating the Density Paradox in Bayesian Active Learning" (DCPV).

These artifacts support the reproducibility of our findings, demonstrating that standard Bayesian Active Learning methods (like Predictive Variance) often fail in Out-of-Distribution (OOD) settings, and that our proposed method (Robust PV) effectively mitigates this issue.

Repository Structure

The repository is organized into three main directories:

1. results/ (Experimental Metrics)

Contains raw .npy files recording the performance metrics for every round of Active Learning (up to 50 rounds).

• Structure: results/{metric}_{method}.npy
• Key Metrics:
  • ood_rate_*.npy: Percentage of selected samples that were OOD (Lower is better).
  • rmse_*.npy: Root Mean Squared Error on the Test Set (Lower is better).
• Methods:
  • random: Random Sampling (Baseline).
  • predictive_variance: Standard PV (Fails in OOD).
  • robust_pv: Our proposed method (Weighted by Mahalanobis density).
  • robust_pv_gmm: Our proposed method (Weighted by GMM density).

2. plots/ (Visualizations)

Generated figures illustrating the "Density Paradox" and the effectiveness of our solution.

• Key Figures:
  • ood_composite_2x4.png: Comprehensive view of OOD selection rates and RMSE across all datasets (MNIST, Fashion-MNIST, CIFAR-10, SVHN).
  • summary_tradeoff_v2.png: Trade-off analysis between OOD robustness and ID performance.

3. data/ (Source Datasets)

Contains the processed datasets used in the experiments to ensure exact reproducibility of splits and preprocessing.

• Datasets:
  • CIFAR-10 / CIFAR-100 (Python version)
  • SVHN (Format 2, 32x32)
  • MNIST / Fashion-MNIST (Raw binaries)

Usage

You can download these artifacts directly or use the huggingface_hub library.

Option 1: Download via Python

from huggingface_hub import snapshot_download

# Download the entire repository
path = snapshot_download(repo_id="darcook/DCPV-Artifacts", repo_type="dataset")
print(f"Artifacts downloaded to: {path}")

Option 2: Load Results with datasets

(Example for listing files)

from huggingface_hub import list_repo_files

files = list_repo_files(repo_id="darcook/DCPV-Artifacts", repo_type="dataset")
print(files)

Citation

If you use these artifacts, please cite our work (Anonymized for Blind Review).