Output details
11 - Computer Science and Informatics
Royal Holloway, University of London
An identity for kernel ridge regression
<24>This paper was selected by the ALT2010 programme committee for a journal special issue. Batch and on-line are the two fundamentally different scenarios of machine learning usually requiring different learning methods or separate theoretical analysis. This paper compares the use of ridge regression (Gaussian processes regression) in these two modes and establishes a general identity connecting the loss of ridge regression in the two modes. The performance of ridge regression in both cases on compact domains is shown to match asymptotically. The paper impacts on practical applications of regression in robotics, for example at the
Italian Institute of Technology, Genoa.