CS: Large Scale Compressive Sensing, Robust PCA for voting in Ensemble algorithms and Spectrum Sensing

After yesterday's entry, Danny Bickson sent me the following:

One paper that is related to compressed sensing is featured here:

http://bickson.blogspot.com/2011/05/large-scale-logistic-regression.html

we are discussing how to implement parallel large scale compressed sensing where there are two feasible approaches: parallel stochastic gradient descent or parallel coordinate descent. We chose to implement the latter option (but we also compare it to the first). The paper is here: http://arxiv.org/abs/1105.5379

We have already featured the paper but you want to read Danny's blog entry.

Another entry following up on the past few days is Bob's blog entry on the voting mechanism to be used in an ensemble approach. Go read it, I'll wait.

In the comment, I mentioned an additional way we could perform this voting between different solvers' results:

Here is another suggestion, we could also put all the solutions as columns of a new matrix called B and look for a factorization of that matrix as: B = A + E + Z , where A is a low rank matrix (rank 1), E is a sparse matrix and Z is a "noisy" matrix ( http://perception.csl.uiuc.edu/matrix-rank/sample_code.html ) or B = A + E (Robust PCA, http://www-stat.stanford.edu/~candes/papers/RobustPCA.pdf ). Of note the comparison between different algorithms: http://perception.csl.uiuc.edu/matrix-rank/sample_code.html#Comparison

Of related interest:

• Recovery of Compressively Sensed Normally-distributed Sparse Signals in Noise
• Paper of the Day (Po'D): An Analysis of Single-Layer Networks in Unsupervised Feature Learning Edition by Pardis Noorzad

Gonzalo Vazquez Vilar is continuing his spectrum sensing summary presented at ICASSP in Cooperative spectrum sensing and resource allocation at ICASSP 2011