AICS is researching the applied use of Artificial Intelligence for several social good applications. Our work has been used to fight human trafficking, visualize and manage information efficiently during crises, detect illicit networks of finance, and quantify gender bias in literature.
We also regularly collaborate with domain-experts to build such systems to ensure responsible use and transparency.
Since the Google Knowledge Graph was publicized in 2011, knowledge graphs (KGs) have become a key component in neurosymbolic AI, with applications ranging from web search to recommendations. Building domain-specific KGs remains a challenging problem, however.
In this interdisciplinary project, we build algorithms, frameworks and tools for semi-automatically constructing domain-specific KGs from raw data. Domains where our methods have been applied include e-commerce, healthcare, and social good applications like human trafficking.
Machine common sense (MCS) is the problem of getting computers to understand common sense -- a wide range of simple facts about people and ordinary life -- and has been recognized by many as a grand challenge in AI since the 1950s. It spans abilities ranging from an intuitive
sense of physics, to the everyday sociology we assume when we talk to other people. Recent progress in large language models and generative AI has led to remarkable progress on MCS, yet there is a lack of consensus on how much common sense today’s AI actually possesses. This project, funded by DARPA, looks at the question both scientifically and from an engineering perspective. Ultimately, we aim to both build systems that have MCS, and to prove that they do.
Open-world learning (OWL) has taken on new importance in recent years as AI systems continue to be applied in real-world environments where structural violations of expectation can occur with non-trivial frequency. Such environmental changes can impact AI performance
profoundly, ranging from overt catastrophic failures to non-robust behaviors that do not take changing context into account. Evaluating OWL algorithms is also a fundamental research area. Funded by DARPA, we aim to build OWL simulation environments (e.g., based on games like Monopoly and Poker) and develop scientific approaches to better understanding the nature of OWL itself.
The broad availability of datasets in complex domains like social media, corporate filings and transportation has led to exciting advances in, and convergence of, research areas like network theory and computational social science.
In this portfolio of projects, we use data and computational methods to study such complex systems. These projects range from developing a better understanding of policy impacts during COVID-19 to quantifying gender bias in literature. Our most recent project aims to study the structure of political campaign finance in the US House of Representatives.
Even before the advent of ChatGPT and other such large language models (LLMs), neural language models like BERT had led to impressive performance on a variety of AI tasks. In this portfolio of projects, we study the properties of such models as 'cognitive machines'.
This is a novel area of research, and we expect regular and significant updates to these projects in the foreseeable future.
