Ever since the 1950’s, the Artificial Intelligence (AI) research community has been obsessed with developing a system that mimics human reasoning abilities. The field has experienced several, sometimes cyclic, paradigm shifts from the symbolic systems in the 50’s to probabilistic reasoning and programming in the 2000’s and modern deep learning in the 2010’s. Despite significant progress, our current state of the art reasoning systems lag behind human capabilities on generalizability, scalability and handling uncertainty. The overarching goal of my research is to build machine learning algorithms for automated and commonsense reasoning.
My work has three main facets for achieving this goal.
- Extrapolation: Data-driven algorithms refer to statistical models that do not need to be explicitly programmed and are instead “trained” on many examples. Often in reasoning the computer encounters examples that are harder than the ones seen during training. I develop novel deep learning and probabilistic learning algorithms that extrapolate to harder test examples.
- Interpretability: Despite the data-driven algorithms’ success, current machine learning models struggle to extract commonsense knowledge from data alone. This is because data contains little information about the commonsense knowledge that went into labeling or annotating it. On the other hand, model-driven algorithms (e.g. rule-based systems) that are programmed for a specific task, explicitly represent commonsense knowledge in terms of interpretable rules. But these models often lack coverage and are susceptible to uncertainty. To use the best of both worlds, I develop novel Neuro-Symbolic learning algorithms, which are hybrid models that leverage the robustness of connectionist methods and the soundness of symbolic reasoning to effectively integrate learning and reasoning.
- Instructability: Recently, the advent of conversational agents, such as Siri and Alexa, has allowed humans to verbally interact with computers. Using these verbal interactions as a means for instructing computers is currently under-explored in machine learning. My research focuses on utilizing human instructions to teach computers
Read more about my research projects here