论文部分内容阅读
稀疏学习由于其简约特性和计算优势而获得了越来越多的关注,在具有稀疏性的条件下,许多计算问题可以在实践中得到有效的处理。而结构化稀疏学习则进一步将结构信息进行编码,在多个研究领域取得成功。随着各类型结构的发现,人们相继提出了各种结构化正则函数。这些正则函数通过利用特定的结构信息极大提高了稀疏学习算法的性能。在本文中,我们从想法、形式化、算法和应用等方面系统的回顾了结构化稀疏学习。我们将这些算法置于最小化损失函数和惩罚函数的统一框架中,总结了算法的开源软件实现,并比较了典型优化算法解决结构化稀疏学习问题时的计算复杂度。在实验中,我们给出了无监督学习在结构化信号恢复和层次化图像重建中的应用,以及具有图结构引导的逻辑回归的在监督学习中的应用。
Due to its simplicity and computational advantages, sparse learning has drawn more and more attention. Under the condition of sparseness, many computational problems can be effectively dealt with in practice. Structured sparse learning further encodes structural information and achieves success in several research areas. With the discovery of various types of structures, many kinds of structured regular functions have been proposed one after another. These regular functions greatly improve the performance of sparse learning algorithms by using specific structural information. In this paper, we systematically review structured sparse learning from the aspects of idea, formalization, algorithm and application. We put these algorithms in the unified framework of minimization of loss function and penalty function, summarize the open source software implementation of the algorithm, and compare the computational complexity of the typical optimization algorithm to solve the structured sparse learning problem. In experiments, we present the application of unsupervised learning in structured signal restoration and hierarchical image reconstruction, and the application of supervised learning with graph-guided logistic regression.