Many source separation algorithms fail to deliver robust performance in presence of artifacts introduced by cross-channel redundancy, non-homogeneous mixing and high-dimensionality of the input signal space. In this paper, we propose a novel framework that overcomes these limitations by integrating learning algorithms directly with the process of signal acquisition and sampling. At the core of the proposed approach is a novel regularized max-min optimization approach that yields "sigma-delta" limit-cycles. An on-line adaptation modulates the limit-cycles to enhance resolution in the signal subspaces containing non-redundant information. Numerical experiments simulating near-singular and non-homogeneous recording conditions demonstrate consistent improvements of the proposed algorithm over a benchmark when applied for independent component analysis (ICA).