PROJECT TITLE :

Efficient Robust Conditional Random Fields - 2015

ABSTRACT:

Conditional random fields (CRFs) are a versatile yet powerful probabilistic approach and have shown benefits for in style applications in numerous areas, including text analysis, bioinformatics, and pc vision. Traditional CRF models, however, are incapable of selecting relevant options also suppressing noise from noisy original features. Moreover, conventional optimization strategies often converge slowly in solving the training procedure of CRFs, and will degrade significantly for tasks with a giant number of samples and features. In this paper, we propose sturdy CRFs (RCRFs) to simultaneously choose relevant features. An optimal gradient technique (OGM) is further designed to coach RCRFs efficiently. Specifically, the proposed RCRFs use the l1 norm of the model parameters to regularize the target utilized by traditional CRFs, therefore enabling discovery of the relevant unary features and pairwise options of CRFs. In each iteration of OGM, the gradient direction is set jointly by this gradient together with the historical gradients, and the Lipschitz constant is leveraged to specify the correct step size. We tend to show that an OGM will tackle the RCRF model training terribly efficiently, achieving the optimal convergence rate O(1/k2) (where k is the amount of iterations). This convergence rate is theoretically superior to the convergence rate O(one/k) of previous first-order optimization strategies. Intensive experiments performed on 3 sensible image segmentation tasks demonstrate the efficacy of OGM in training our proposed RCRFs.