Independent component analysis is a classical multivariate tool used for estimating independent sources among collections of mixed signals. However, modern forms of data are typically too complex for the basic theory to adequately handle. In this thesis extensions of independent component analysis to three cases of non-standard data structures are developed: noisy multivariate data, tensor-valued data and multivariate functional data. 

In each case we define the corresponding independent component model along with the related assumptions and implications. The proposed estimators are mostly based on the use of kurtosis and its analogues for the considered structures, resulting into functionals of rather unified form, regardless of the type of the data. We prove the Fisher consistencies of the estimators and particular weight is given to their limiting distributions, using which comparisons between the methods are also made.