The target of this research initiative is the 'control problem' in A.I. safety: how to create superintelligent systems that could be relied upon to preserve what we value. Projects within its scope include the safe application of superintelligent question-answering systems ('oracles') and methods for designing 'corrigible' superintelligent agents--systems whose behaviors could be corrected at run-time.
Aristo is a system that answers questions from standardized science exams at multiple grade levels. The process involves three basic functions: (1) an ability to acquire and process facts into a structured knowledge base; (2) an ability to 'understand' exam questions and analyze their accompanying diagrams; and (3) an ability to answer questions using entailment, statistical analysis, and inference methods.
DeepQA was a project by IBM to develop a software architecture that could answer questions over a wide range of subjects in human rather than computer terms. Watson, the product of DeepQA, uses machine learning and natural language processing to answer questions, quickly extract information, and analyze large amounts of unstructured data. Recent applications of Watson stretching from healthcare to customer service have been upshots of DeepQA technology. In 2011, Watson defeated Jeopardy champions Ken Jennings and Brad Rutter to win the 1 million dollar Jeopardy prize.
Euclid is a project to develop systems that solve math and geometry problems. Recently, it developed GeoS, an automated system that solves high school geometry problems by interpreting question diagrams and text in natural language. The system achieved a 49% score on official SAT questions and a 61% on practice SAT questions.
This project concerns the acquisition, learning, and representation of visual knowledge. It seeks to build systems that exhibit visual common-sense and are able to integrate visual information with textual information to improve knowledge acquisition and reasoning. Problems it has tackled or is tackling are the prediction of the dynamics of objects in static images (Newtonian Neural Network), the visual learning of concepts such as 'chair' (Levan), the visual verification of relation phrases such as 'does the dog eat ice cream' (VisKE), and the visual recognition of daily human activities (Charades).