Date(s) - 08/11/2019
14 h 00 min - 15 h 00 min
Catégories Pas de Catégories
Juho ROUSU (Aalto University, Helsinki)
Canonical correlation analysis (CCA) methods find multivariate relations in two-view data settings. CCA can be seen as a relative of principal component analysis, when the objective is to explain covariance between two views rather than variance within one view. When applying CCA methods, the interest is typically to find the related variables in the two views, and to uncover the relation that couples the variables. In general, it is difficult to achieve these objectives if the underlying multivariate relation is non-linear and the data is high-dimensional. CCA methods in the literature tend to be either non-linear (such as kernel CCA and deep CCA) or sparse (sparse CCA) but not both. In this presentation, I discuss recent progress in learning CCA models where the underlying relations are both non-linear and sparse. Our new gradKCCA method is based on mapping a sparse projection of the data through a non-linear kernel function. From an another angle, we can view the method as solving an ‘optimal preimage’ problem for KCCA. The model is optimized using an alternating projected gradient algorithm. Unlike KCCA, the method does not rely on a kernel matrix, and hence scales up effortless to large datasets. Experimentally, gradKCCA shows competitive predictive performance to KCCA and deep CCA models, while being orders of magnitude faster to optimize.
Uurtio, V., Bhadra, S. and Rousu, J., 2019, Large-scale sparse kernel canonical correlation analysis, ICML’2019
Juho Rousu is a Professor of Computer Science at Aalto University, Finland. Rousu obtained his PhD in 2001 form University of Helsinki, while working at VTT Technical Centre of Finland. In 2003-2005 he was a Marie Curie Fellow at Royal Holloway University of London. In 2005-2011 he held Lecturer and Professor positions at University of Helsinki, before moving to Aalto University in 2012 where he leads a research group on Kernel Methods, Pattern Analysis and Computational Metabolomics (KEPACO). Rousu’s main research interest is in learning with multiple and structured targets, multiple views and ensembles, with methodological emphasis in regularised learning, kernels and sparsity, as well as efficient convex/non-convex optimisation methods. His applications of interest include metabolomics, biomedicine, pharmacology and synthetic biology.