Statistical Modeling and Inference for unsupervised Learning at largE-Scale – SMILES

- latent variable models for high-dimensional regression, including functional regression and functional regressions mixture models
- latent variable models for high-dimensional classification and clustering in high-dimensional scenarios
- latent variable models for unsupervised learning and bioacoustics

- theoretical, methodological and computational results on approximation, estimation and model selection capabilities in latent variable models, in particular for mixtures and mixtures of experts

deep latent variable models and distributed mixtures for clustering

see the mid-term report

Large-scale data analysis is an inherently multidisciplinary area and is becoming of broader interest for today's society. SMILES is a collaborative fundamental research project that aims at introducing an unsupervised statistical modeling framework and scaled inference algorithms for transforming large-scale data into knowledge. It considers the large-scale context as a whole, with its main issues related to inference from a big volume of data of very high dimension and underlying complex hidden structures. The key tenet of SMILES is to introduce large-scale regression-based sparse (non)parametric models for data representation, and large-scale latent data models for unsupervised data classification. The knowledge extraction will namely consist in automatically retrieving hidden structures, summarizing prototypes, groups, sparse representations. We consider different data settings, including functional data, multimodal bioacoustical data, and biological data.