Canonical correlation analysis and Transfer Learning

Many domain adaptation algorithms are based on finding a shared projection of the source and target data such that they both start to look as if being drawn from the same distribution. So yes, it's possible to use CCA for domain adaptation as, on a high level, CCA extracts features from two data sources such that the projected features of the two sources becomes maximally correlated. A recent example where CCA was used for domain adaptation is the zero-shot learning for domain adaptation paper that uses novel target specific features based on how these features relate to certain shared features (present in both domains) that can be extracted using CCA. You should also look at John Blitzer's thesis where a connection between the structural correspondence learning (a domain adaptation technique) and CCA has been shown. CCA has also been used for cross-lingual classification (e.g., this paper) in NLP (and various other cross-lingual IR applications).

CCA can likewise also be used for multitask learning. E.g. see the paper Integrating Multiple-Study Multiple-Subject fMRI Datasets Using Canonical Correlation Analysis (though it's actually a special case of multitask learning -- predicting multivariate responses or multiple labels -- so each response/label can be thought of as being as "task").

Another interesting, though a bit different from domain adaptation and multitask learning, is the case of multi-view learning where data consists of multiple views. CCA can be applied in these settings too (e.g., see the multiview clustering via Canonical Correlation Analysis paper where CCA has been used for clustering using multiple views of the data).