Kernel signal-to-noise ratio for machine learning
Place: Large Lecture Room
Affiliation: Image and Signal Processing Group (ISP), Universitat de València, Spain
In this talk I will introduce a powerful kernel function for machine learning and signal processing. The proposed kernel signal to noise ratio (KSNR) maximizes the signal variance while minimizes the estimated noise variance in a reproducing kernel Hilbert space (RKHS). The KSNR can be used in any kernel method to deal with correlated (possibly non-Gaussian) noise. We illustrate the method in nonlinear regression examples, dependence estimation and causal inference, nonlinear classification in channel equalization, and nonlinear feature extraction from high-dimensional satellite images. The family of methods generalize kernel PCA, least squares SVM, and kernel ridge regression in the case of i.i.d. noise. Experimental results show that the proposed KSNR yields more fitted solutions and extracts more noise-free features when confronted with standard approaches.