Research
We build practical, theory-grounded interaction-based explanation methods to understand what deep neural networks learn and how they learn it. Our work treats feature interactions as a unifying language for interpretability, connecting attribution, learning dynamics, generalization, and robustness. By combining rigorous interaction theory with scalable monitoring and real-world validation, we help teams explain, audit and debug models.
Attribution
We develop principled attribution methods that explain predictions by allocating credit to inputs and their combinations, unifying majority post-hoc explainers and improving Shapley-style efficiency and faithfulness.
Attribution
We develop principled attribution methods that explain predictions by allocating credit to inputs and their combinations, unifying majority post-hoc explainers and improving Shapley-style efficiency and faithfulness.
Interaction Theory
We formalize and extract interaction primitives, which are sparse, reusable building blocks that capture compositional concepts, providing a compact, interpretable view of how networks represent knowledge.
Interaction Theory
We formalize and extract interaction primitives, which are sparse, reusable building blocks that capture compositional concepts, providing a compact, interpretable view of how networks represent knowledge.
Robustness
We analyze why models fail under adversarial pressure, distribution shift, or fine-tuning, and propose tools to assess and improve stability of learned concepts and fingerprints.
Robustness
We analyze why models fail under adversarial pressure, distribution shift, or fine-tuning, and propose tools to assess and improve stability of learned concepts and fingerprints.
Robustness
We analyze why models fail under adversarial pressure, distribution shift, or fine-tuning, and propose tools to assess and improve stability of learned concepts and fingerprints.
Applications
We translate these ideas into practice across vision and 3D understanding, privacy/attribute obfuscation, and LLM behavior auditing (e.g., legal reasoning), demonstrating impact in real deployment scenarios.
Applications
We translate these ideas into practice across vision and 3D understanding, privacy/attribute obfuscation, and LLM behavior auditing (e.g., legal reasoning), demonstrating impact in real deployment scenarios.
Applications
We translate these ideas into practice across vision and 3D understanding, privacy/attribute obfuscation, and LLM behavior auditing (e.g., legal reasoning), demonstrating impact in real deployment scenarios.
