Resource-bounded adaptive inference of accurate conditional probability estimates from data

Effective use of machine learning to make predictions is critical to many advanced interactive systems. Such systems often operate under strict time and memory constraints. In consequence they use very fast learning algorithms that require little memory but deliver less reliable predictions than more resource intensive approaches. This project will deliver a new generation of learning algorithms that adapt to prevailing computational constraints, using such time and memory as is available to deliver the best predictions possible within those constraints. This will exploit computational resources that would otherwise be left idle, thereby improving the performance of many advanced interactive applications.