research.html

---
layout: page
title: Research
permalink: /research/
---
<strong>Unrolling FISTA to solve Low-Rank related problems</a></strong> <br><br>
<p style="text-align:justify"><i>Abstract: </i>Based on the avaiable of BP Algorithm for SVD operator, it is possible to unroll ISTA into a deep nueral network for low-rank regularization problems. To accelerate the convergence rate while avoiding the design of the momentum term manually, we come up with a structure similar to LSTM aiming at learning the momentum term adaptively. This method also gives an inspiration to unroll RPCA into deep neural network by disentangling RPCA into low-rank problem and sparse problem.</p>
<br>
<strong>Laplace K-log: A new Laplace regularized robust algorithm for clustering </a></strong><br><br>
<p style="text-align:justify"><i>Abstract: </i>K-means is a common stretegy for clustering, owing to it's efficient computation and guarantee convergency. Nonetheless, using L2 norm as measurement for distance makes K-means sensitive to outliers. K-mediods is based on L1 norm thus is more robust, however, it makes the algorithm more computational expensive. We put forward K-log, which combine the high efficiency of K-means and robustness of K-medoids. Moreover, to entangle the spatial consistency into the algorithm, a Laplace regularizer is added and Laplace K-log is proposed.  </p><br>