Inference and learning in the Hopfield model
In this summary we study reconstruction of Hopfield networks in memory phase from experimental data. Nouergodiclty of whole configuration space in Hopfield networks makes reconstruction extremely hard for fast mean-field inverse algorithms like independent pair approximation method or inverse of TAP equations approach. We suggest methods based on message passing algorithm to do reconstruction when experimental data comes from one or few basins of attractions. The idea is to exploit information of other patterns hidden in cross talk noise of experimental data of one basin. We extend this idea to an unsupervised learning algorithm in which patterns are weakly presented to system as external fields. The algorithm learns from each field both condensed pattern(from signal) and other patterns(from crosstalk noise). In this way system will learn a large number of patterns without entering into spin glass phase, as it happens in Hebbian Learning.
Pan Zhang Abolfazl Ramezanpour Riccanlo Zeccbina
Polileenico di Torino, C.so Duca degli Ahruzzi 24. 1-10129 Torino. Italy
国际会议
International Workshop on Statistical Physics and Computer Sciences(统计物理与计算机科学交叉研究国际研讨会 )
北京
英文
286-293
2010-07-08（万方平台首次上网日期，不代表论文的发表时间）