Large pre-trained language models consistently top the leaderboard for benchmarks across a variety of NLP tasks. However, due to the high complexity of these models, it is challenging for researchers, developers, and AI end-users to understand how these models make prediction decisions. Understanding model behavior is crucial for identifying the limitations of AI systems, diagnosing flaws in model architectures and datasets, and building trust with the global community who are affected by AI algorithms in their daily lives. In this work, we leverage the Tiered Reasoning for Intuitive Physics dataset and framework to investigate model behavior on the task of physical commonsense reasoning. Using an explainability tool, LIME, and a prompting approach, we conduct analysis to uncover model behavior in story classification. Our findings identify several factors that affect model decisions and raise some topics for consideration when designing a training pipeline and benchmarks for the task of commonsense. Paper

Built a system to evaluate the spread of covid-19 in indoor environments, by simulating human behavior using Reinforcement Learning. Through realistic models of home environments, and human behavior, we are understanding the spread of covid-19 due to human movements, to identify possible strategies to mitigate such spread. Click here for a presentation video.