By Maryam Mahdaviani (based on joint work with Nando de Freitas)
Previous studies have demonstrated the accuracy of semi-supervised and active learning with Gaussian fields. However, these algorithms are computationally expensive. In order to apply these ideas in interactive robotics, a reduction in cost is essential. Taking advantage of sophisticated numerical algorithms, we propose a solution to resolve this obstacle. Considerable speed ups are achieved using Krylov subspace methods and the Fast Gauss Transform. The fast algorithms have been implemented for a simple recognition task on Sony’s ERS-7 Aibo. As a result of this work, LCI’s Aibo has become an AiboB, a robot dog with a Brain who can be trained to recognize household objects!