My research focuses on the computational foundations of intelligent behavior. We develop theories and systems pertaining to intelligent behavior using a unified methodology -- at the heart of which is the idea that learning plays a central role in intelligence.
Within this broader program, I study machine learning and inference methods that enable natural language understanding. My work addresses fundamental questions about learning, reasoning, and interaction, and applies these foundations to a wide range of natural language processing (NLP) problems. My theoretical work on the Learning to Reason framework laid the foundations for integrating learning into intelligent decision systems and gave rise to a highly influential constrained optimization framework that augments statistical machine learning with declarative constraints (ILP for NLP) -- establishing the foundations of Neuro-Symbolic AI.
We introduced Zero-shot Text Classification to the AI community (originally named "Dataless Classification"), pioneering semantic embeddings and label-aware models that are now foundational to modern LLMs. We also pioneered self-supervised learning paradigms -- starting with the context-sensitive spelling correction work in the late 1990s -- long preceding the self-training approaches central to today's large language models. Within these frameworks we have developed state-of-the-art solutions for semantic role labeling, co-reference resolution, textual entailment, named entity recognition, and entity-linking. Much of this, and later work, emphasizes incidental supervision to overcome the difficulty of supervising complex problems, and addresses the trustworthiness and reliability of LLM reasoning through Neuro-Symbolic approaches.
I have also worked on fundamental problems in probabilistic reasoning, temporal and event-driven reasoning, and the trustworthiness of information -- anticipating current concerns about misinformation and LLM hallucination.
In computer vision and multimodal AI, I pioneered early discriminative part-based methods for object recognition and have recently shown that genuine visual understanding requires reasoning that goes beyond raw pixel recognition.
In industrial leadership roles at AWS and Oracle, I have made foundational contributions to Generative AI at industry scale: leading the scientific effort behind Amazon's first-generation GenAI products -- Amazon Titan, Amazon Bedrock, and Amazon Q -- and contributing to innovations spanning coding agents, reasoning and planning agents, and Responsible AI. At Oracle, I founded and lead the GenAI scientific effort and, among multiple efforts, am leading work on retrieval and reasoning over structured and unstructured data, with demonstrated leadership in NL-to-SQL and related tasks.
For more details, check the left panel for Research Contributions, short bio, or CV.
