Knowledge Injection XAI

Goal

To develop a distributed framework that can predict out-of-distribution (OOD) robustness of models using only Explainable AI (XAI) metrics extracted from clean data, eliminating the need for corrupted test sets during evaluation.

Technical Approach

• Built a Spark pipeline integrating DINOv2 as the backbone vision encoder with LoRA (Low-Rank Adaptation) fine-tuning and XGBoost for robustness forecasting.
• Extracted XAI metrics from clean data to predict how models would perform under distribution shifts and corruptions.
• Compared low-rank vs. high-rank adapter configurations to analyze their impact on corruption resistance.

Key Metrics & Results

• Achieved 0.739 ROC-AUC on the OOD robustness prediction task.
• Demonstrated that low-rank adapters resist corruptions 19% better than high-rank ones, providing actionable insights for model design.
• The distributed Spark pipeline enables scalable evaluation across large model collections.

Tech Stack

• Python, PyTorch, Apache Spark, DINOv2, LoRA, XGBoost