Deep Learning Theory & Methods

My research in deep learning foundations develops training algorithms with provable convergence guarantees, distributed learning methods for large-scale systems, and techniques for handling challenging data scenarios.

Direct Error-Driven Learning: Novel training algorithms that directly minimize classification error rather than surrogate losses, providing theoretical guarantees for deep network training on high-dimensional data. This work bridges optimization theory with practical deep learning.

Distributed Learning: Game-theoretic and min-max optimization frameworks for training neural networks across distributed computing environments, with applications to big data classification problems where data cannot be centralized.

Domain Adaptation: Addressing distribution shift between training and deployment through adversarial and game-theoretic approaches that learn domain-invariant representations, critical for deploying ML models in evolving real-world environments.

Dimension Reduction: Hierarchical and multi-step nonlinear methods for reducing dimensionality of big data while preserving discriminative information for downstream classification and regression tasks.

Privacy-Preserving ML: Federated learning approaches for scientific applications where data cannot be centralized due to privacy, security, or governance constraints across institutions and national laboratories.

Multi-Agent Systems: Credit assignment methods for multi-LLM agent systems, addressing the challenge of attributing rewards and blame when multiple AI agents collaborate on complex tasks.

Publications