Neural networks for efficient nonlinear online clustering

Unsupervised learning techniques, such as clustering and sparse coding, have been adapted for use with data sets exhibiting nonlinear relationships through the use of kernel machines. These techniques often require an explicit computation of the kernel matrix, which becomes expensive as the number of inputs grows, making it unsuitable for efficient online learning. This paper proposes an algorithm and a neural architecture for online approximated nonlinear kernel clustering using any shift-invariant kernel. The novel model outperforms traditional low-rank kernel approximation based clustering methods, it also requires significantly lower memory requirements than those of popular kernel k-means while showing competitive performance on large data sets.