Topology-Based Active Learning

A common problem in simulation and experimental research involves obtaining time-consuming, expensive, or potentially hazardous samples from an arbitrary dimension parameter space. For example, many simulations modeled on supercomputers can take days or weeks to complete, so it is imperative to select samples in the most informative and interesting areas of the parameter space. In such environments, maximizing the potential gain of information is achieved through active learning (adaptive sampling). Though the topic of active learning is well-studied, this paper provides a new perspective on the problem. We consider topologybased batch selection strategies for active learning which are ideal for environments where parallel or concurrent experiments are able to be run, yet each has a heavy cost. These strategies utilize concepts derived from computational topology to choose a collection of locally distinct, optimal samples before updating the surrogate model. We demonstrate through experiments using a several different batch sizes that topology-based strategies have comparable and sometimes superior performance, compared to conventional approaches. Topology-Based Active Learning Dan Maljovec School of Computing University of Utah Salt Lake City, UT 84112 [email protected] Bei Wang Scientific Computing and Imaging Institute University of Utah Salt Lake City, UT 84112 [email protected] John Moeller School of Computing University of Utah Salt Lake City, UT 84112 [email protected] Valerio Pascucci Scientific Computing and Imaging Institute University of Utah Salt Lake City, UT 84112 [email protected]