AAAI Publications, The Twenty-Sixth International FLAIRS Conference

Font Size: 
Novelty Detection Using Sparse Online Gaussian Processes for Visual Object Recognition
Ruben Ramirez-Padron, Boris Mederos, Avelino J. Gonzalez

Last modified: 2013-05-19


Gaussian processes (GPs) have been shown to be highly effective for novelty detection through the use of different membership scores. However, applications of GPs to novelty detection have been limited only to batch GP, which require all training data at once and have quadratic space complexity and cubic time complexity. This paper proposes the use of sparse online GP (SOGP) for novelty detection, overcoming these limitations. Our experiments show that SOGP-based novelty detection is capable of achieving performances similar to those from batch GP, even under strong sparseness constraints. Additionally, it is suggested here that membership scores that combine the posterior mean and the posterior variance of the GP might be better fitted to novelty detection than scores leveraging only one of the two posterior moments.


outlier detection; novelty detection; Gaussian process; sparse online learning; visual object recognition

Full Text: PDF