Contributor(s)CALIFORNIA UNIV REGENTS OAKLAND
KeywordsStatistics and Probability
Full recordShow full item record
AbstractDespite great advances in artificial intelligence (AI) research over the last fifty years, computers are still far worse than people at solving many important problems, such as learning language, inferring categories of objects from just a few examples, and identifying causal relationships. The goal of this project was to develop automated systems that can match human performance in problems of this kind. The approach that was taken to achieving this goal is one that contributed to the first AI systems: identifying the formal principles that characterize how people solve these problems. This required combining mathematical tools from computer science and statistics with the empirical methods of cognitive psychology. By exploiting the interplay between these disciplines, the resulting research provided insight into how we can make machines learn, and a deeper understanding of how the human mind works. Building on recent work in both AI and cognitive science, this project explored the possibility that Bayesian statistics can provide formal solutions to inductive problems. Bayesian statistics is based upon a simple principle that dictates how a rational agent should change his or her beliefs in light of evidence, called Bayes' rule. Bayes' rule is a principled way to combine constraints on hypotheses from prior knowledge with the evidence provided by data, and motivates much contemporary research in statistical artificial intelligence and machine learning. The research was divided into two objectives, each addressed by three different lines of work. The remainder of this report summarizes the results of these lines of work. Publications resulting from the work are cited and appear in the reference list. These publications summarize the data collected as part of the grant and describe relevant models at a level where they can be re-implemented.