Source separation in the presence of side-information
Zahra Sabetsarvestani · UCL Discovery (University College London) · 2020
The source separation problem involves the separation of unknown signals from their mixture. This problem is relevant in a wide range of applications from audio signal processing, communication, biomedical signal processing and art investigation to name a few. There is a vast literature on this problem which is based on either making strong assumption on the source signals or availability of additional data. This thesis proposes new algorithms for source separation with side information where one observes the linear superposition of two source signals plus two additional signals that are correlated with the mixed ones. The first algorithm is based on two ingredients: first, we learn a Gaussian mixture model (GMM) for the joint distribution of a source signal and the corresponding correlated side information signal; second, we separate the signals using standard computationally efficient conditional mean estimators. This also puts forth new recovery guarantees for this source separation algorithm. In particular, under the assumption that the signals can be perfectly described by a GMM model, we characterize necessary and sufficient conditions for reliable source separation in the asymptotic regime of low-noise as a function of the geometry of the underlying signals and their interaction. It is shown that if the subspaces spanned by the innovation components of the source signals with respect to the side information signals have zero intersection, provided that we observe a certain number of linear measurements from the mixture, then we can reliably separate the sources; otherwise we cannot. The second algorithms is based on deep learning where we introduce a novel self-supervised algorithm for the source separation problem. Source separation is intrinsically unsupervised and the lack of training data makes it a difficult task for artificial intelligence to solve. The proposed framework takes advantage of the available data and delivers near perfect separation results in real data scenarios. Our proposed frameworks – which provide new ways to incorporate side information to aid the solution of the source separation problem – are also employed in a real-world art investigation application involving the separation of mixtures of X-Ray images. The simulation results showcase the superiority of our algorithm against other state-of-the-art algorithms.