KERMES - Advances in Kernel methods for Structured Data

KERMES is a NoE on kernel methods for structured data, funded by the Spanish Ministry of Economy and Competitiveness, TEC2016-81900-REDT running between 01/17 and 12/18. In the last decade there has been an increasing availability of structured data coming from different sensory devices, with different complexities and noise sources: from time series of geospatial data and remotely sensed images, to biosignals and medical records, or Internet and communication data streams. All of them share and exhibit common traits: highly redundant spatial-temporal features, complex multimodal feature relations, uneven sampling and heterogeneous noise sources. The analysis of structured data is challenging, and traditional machine learning methods work under the assumption that signal samples are independent and identically distributed (i.i.d.).

Kernel machines and related Bayesian approaches, such as Gaussian processes, have been widely and successfully used in practice for dealing with such data structures for regression and classification. Kernel methods allow to encode prior knowledge about the data characteristics, to learn the underlying latent functions explaining the data; and allow the combination of different data modalities. The long-term vision of KERMES is tied to open new frontiers and foster research towards new kernel algorithms, a stepping stone before the more ambitious far-end goal of machine reasoning. We aim to learn data structures, to scale methods to work with millions of samples, derive sensible confidence intervals for the predictions, and to advance in causal inference from empirical data. The members of this network are leading experts in kernel methods.

KERMES will promote multidisciplinary collaboration among researchers in the fields of machine learning, signal processing and Bayesian statistics. We aim to create synergies, organize discussion meetings, publish relevant papers, sponsor researcher's mobility, and create a critical mass that allows to launch future scientific projects. Technology transfer activities will be directly performed in collaboration with companies in this sector as well.