Research Vision

The rapid advancement of foundation models and deep learning has transformed artificial intelligence into a core technology for scientific discovery, healthcare, cybersecurity, autonomous systems, agriculture, and many other safety-critical applications. These models have achieved remarkable performance across diverse tasks, yet their increasing scale and complexity have introduced significant challenges to their trustworthy deployment. Modern AI systems often operate as black boxes, making their internal reasoning difficult to interpret, while simultaneously exhibiting hallucinations, adversarial vulnerabilities, and unpredictable behaviors under distribution shifts. These limitations reduce user trust and hinder the adoption of AI in high-consequence decision-making environments where transparency, reliability, and security are essential.

My research vision is to establish a unified scientific foundation for trustworthy artificial intelligence by bridging AI explainability, robustness, and security. Rather than treating interpretability as a tool for explaining decisions after they are made, I seek to develop intrinsically interpretable learning algorithms whose internal representations are transparent, verifiable, and inherently resistant to adversarial manipulation and hallucinations. My long-term goal is to enable AI systems that can reason reliably, justify their decisions, and operate safely in real-world environments.

To realize this vision, my research spans several interconnected directions. The first focuses on developing interpretable neural architectures that incorporate structured representations, prototype learning, concept-based reasoning, and mechanistic interpretability to reveal the internal computation of deep learning models. The second investigates the theoretical relationship between explainability and adversarial robustness by designing learning algorithms that jointly improve model transparency, robustness, and generalization. The third develops trustworthy foundation models capable of mitigating hallucinations, improving reasoning reliability, and defending against adversarial and jailbreak attacks through interpretable internal reasoning mechanisms. Finally, I apply these methodologies to high-impact scientific and societal domains, including medical image analysis, materials discovery, precision agriculture, and scientific machine learning, where trustworthy AI is essential for informed decision-making.