PASCAL - Pattern Analysis, Statistical Modelling and Computational Learning

Robust sparse recovery with non-negativity constraints
Matthias Hein
Workshop on Signal Processing with Adaptive Sparse Structured Representations (SPARS) 2011.


It has been established recently that sparse non-negative signals can be recovered using non-negativity constraints only. This result is obtained within an idealized setting of exact sparsity and absence of noise. We propose non-negative least squares − without any regularization − followed by thresholding for the noisy case. We develop conditions under which one can prove a finite sample result for support recovery and tackle the case of an approximately sparse target. Under weaker conditions, we show that non-negative least squares is consistent for prediction. As illustration, we present a feature extraction problem from Proteomics.

EPrint Type:Article
Project Keyword:Project Keyword UNSPECIFIED
Subjects:Learning/Statistics & Optimisation
ID Code:8706
Deposited By:Matthias Hein
Deposited On:21 February 2012