Phase Transitions and Sample Complexity in Bayes-Optimal Matrix Factorization

被引:75
作者
Kabashima, Yoshiyuki [1 ]
Krzakala, Florent [2 ,3 ]
Mezard, Marc [4 ]
Sakata, Ayaka [5 ,6 ]
Zdeborova, Lenka [7 ,8 ]
机构
[1] Tokyo Inst Technol, Dept Math & Comp Sci, Yokohama, Kanagawa 2268502, Japan
[2] PSL Res Univ, Ecole Normale Super, Lab Phys Stat, F-75005 Paris, France
[3] Sorbonne Univ, Univ Paris 06, F-75005 Paris, France
[4] PSL Res Univ, Ecole Normale Super, Dept Phys, F-75005 Paris, France
[5] Inst Stat Math, Tachikawa, Tokyo 1908562, Japan
[6] Grad Univ Adv Sci SOKENDAI, Hayama, Kanagawa 2400193, Japan
[7] Commissariat Energie Atom Saclay, F-91191 Gif Sur Yvette, France
[8] CNRS, Inst Phys Theor, F-91191 Gif Sur Yvette, France
基金
欧洲研究理事会; 日本学术振兴会;
关键词
Statistical inference; probabilistic matrix factorization; dictionary learning; message passing algorithms; phase transitions; sparse coding; statistical physics; computational barriers; statistical and computational tradeoff; BLIND SOURCE SEPARATION; OVERCOMPLETE DICTIONARIES; SPARSE; COMPLETION; ALGORITHMS; GRAPHS; MODEL; CDMA;
D O I
10.1109/TIT.2016.2556702
中图分类号
TP [自动化技术、计算机技术];
学科分类号
0812 ;
摘要
We analyze the matrix factorization problem. Given a noisy measurement of a product of two matrices, the problem is to estimate back the original matrices. It arises in many applications, such as dictionary learning, blind matrix calibration, sparse principal component analysis, blind source separation, low rank matrix completion, robust principal component analysis, or factor analysis. It is also important in machine learning: unsupervised representation learning can often be studied through matrix factorization. We use the tools of statistical mechanics-the cavity and replica methods-to analyze the achievability and computational tractability of the inference problems in the setting of Bayes-optimal inference, which amounts to assuming that the two matrices have random-independent elements generated from some known distribution, and this information is available to the inference algorithm. In this setting, we compute the minimal mean-squared-error achievable, in principle, in any computational time, and the error that can be achieved by an efficient approximate message passing algorithm. The computation is based on the asymptotic state-evolution analysis of the algorithm. The performance that our analysis predicts, both in terms of the achieved mean-squared-error and in terms of sample complexity, is extremely promising and motivating for a further development of the algorithm.
引用
收藏
页码:4228 / 4265
页数:38
相关论文
共 74 条
[1]   On the uniqueness of overcomplete dictionaries, and a practical way to retrieve them [J].
Aharon, Michal ;
Elad, Michael ;
Bruckstein, Alfred M. .
LINEAR ALGEBRA AND ITS APPLICATIONS, 2006, 416 (01) :48-67
[2]   K-SVD: An algorithm for designing overcomplete dictionaries for sparse representation [J].
Aharon, Michal ;
Elad, Michael ;
Bruckstein, Alfred .
IEEE TRANSACTIONS ON SIGNAL PROCESSING, 2006, 54 (11) :4311-4322
[3]   The ζ(2) limit in the random assignment problem [J].
Aldous, DJ .
RANDOM STRUCTURES & ALGORITHMS, 2001, 18 (04) :381-418
[4]  
[Anonymous], SAMPLE COMPLEXITY DI
[5]  
[Anonymous], 2003, EXPLORING ARTIFICIAL
[6]  
[Anonymous], P IEEE INT S INF THE
[7]  
[Anonymous], 2006, P 2006 ESANN BRUG BE
[8]  
[Anonymous], 2014, Information-theoretically optimal sparse pca
[9]  
[Anonymous], SPARSE LOW IN PRESS
[10]  
[Anonymous], MMSE PROBABILISTIC L