The project objectives as listed in the original proposal are the following:

  1. Development of efficient time-adaptive and sparsity-aware parameter learning schemes for single sensor measurements. In this context, efficiency will be treated with respect to both low error floors and strict complexity guarantees, in order the schemes to be appropriate for real time operation.
  2. Development of systematic mechanisms for the incorporation of generalized notions of sparsity such as block-sparsity, tree-structured sparsity and hierarchically-structured sparsity.
  3. Development of online sparsity-aware learning schemes suitable for multiple-sensor measurements. This objective evolve along two directions. First, to advance the Adaptive projection learning filtering (APLF) methodology with techniques that exploit the correlations in the sparsity patterns between signals, which are acquired with multiple-sensor devises/topologies. This will allow the online sparsity-aware processing of multiple measurement vectors (MMV) for the first time. Second, to extend such schemes and making them appropriate for distributed processing required in WSN.
  4. Evaluation of the developed techniques. The point of this objective is to evaluate the developed algorithms in carefully designed simulations capable of highlighting their pros and cons and offering a platform for fair comparisons with competitive batch and online sparsity-aware algorithms and 5) Implementation of a wireless electrocardiogram (ECG) monitoring system. The ability of the developed techniques to operate in a real-world application will be assessed. This involves the joint recovery of the signals, resulted from a 12-Lead ECG, based on a reduced number of wirelessly transmitted measurements. Signal measurement, transmission and recovery will all be realized in real-time.

Additional objectives adopted during the project:

  1. to extend the sparsity promotion mechanisms associated with the APLF framework to the case of online robust subspace tracking and online distributed dictionary learning,
  2. to adopt a medical application more suitable that the ECG for analysis based on constrained matrix factorization and dictionary learning. For this reason analysis of functional Magnetic Resonance Imaging (fMRI) was adopted
  3. to develop complexity reduction techniques for the online sparsity aware matrix factorization schemes discussed before. That was necessary since the associated approaches followed in the regression/filtering setting specified in the original objectives of the project where not applicable in the matrix factorization setup. For these reasons the approach followed was that of randomized dimensionality reduction.