We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Physics > Data Analysis, Statistics and Probability

Title: Information Perspective to Probabilistic Modeling: Boltzmann Machines versus Born Machines

Abstract: We compare and contrast the statistical physics and quantum physics inspired approaches for unsupervised generative modeling of classical data. The two approaches represent probabilities of observed data using energy-based models and quantum states respectively.Classical and quantum information patterns of the target datasets therefore provide principled guidelines for structural design and learning in these two approaches. Taking the restricted Boltzmann machines (RBM) as an example, we analyze the information theoretical bounds of the two approaches. We verify our reasonings by comparing the performance of RBMs of various architectures on the standard MNIST datasets.
Comments: 7 pages, 4 figures
Subjects: Data Analysis, Statistics and Probability (physics.data-an); Statistical Mechanics (cond-mat.stat-mech); Quantum Physics (quant-ph); Machine Learning (stat.ML)
Journal reference: Entropy 2018, 20(8), 583
DOI: 10.3390/e20080583
Cite as: arXiv:1712.04144 [physics.data-an]
  (or arXiv:1712.04144v1 [physics.data-an] for this version)

Submission history

From: Song Cheng [view email]
[v1] Tue, 12 Dec 2017 06:34:10 GMT (485kb,D)

Link back to: arXiv, form interface, contact.