Procedural Adversarial Learning

TL;DR

The paper introduces Procedural Adversarial Learning (PAL), a framework that uses zeroth-order optimization to discover challenging input instances by exploring procedural generators. Simple-PAT, the proposed algorithm, generates adversarial training samples on demand, improving model robustness across multiple vision tasks. PAL-Dojo provides efficient procedural models for surface normal estimation, segmentation, and pose estimation.

Why It Matters

This approach enables more robust neural networks by proactively identifying failure cases in open input spaces, rather than relying on fixed datasets. It enhances model reliability in real-world scenarios where unexpected inputs are common, especially in safety-critical applications.

Key Findings

• Simple-PAT uses zeroth-order optimization to search procedural generator parameter spaces and generate adversarial training samples dynamically.