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

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

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

Computer Science > Machine Learning

Title: PSD Representations for Effective Probability Models

Abstract: Finding a good way to model probability densities is key to probabilistic inference. An ideal model should be able to concisely approximate any probability while being also compatible with two main operations: multiplications of two models (product rule) and marginalization with respect to a subset of the random variables (sum rule). In this work, we show that a recently proposed class of positive semi-definite (PSD) models for non-negative functions is particularly suited to this end. In particular, we characterize both approximation and generalization capabilities of PSD models, showing that they enjoy strong theoretical guarantees. Moreover, we show that we can perform efficiently both sum and product rule in closed form via matrix operations, enjoying the same versatility of mixture models. Our results open the way to applications of PSD models to density estimation, decision theory and inference.
Comments: 50 pages, 1 table
Subjects: Machine Learning (cs.LG); Statistics Theory (math.ST); Machine Learning (stat.ML)
Cite as: arXiv:2106.16116 [cs.LG]
  (or arXiv:2106.16116v3 [cs.LG] for this version)

Submission history

From: Carlo Ciliberto [view email]
[v1] Wed, 30 Jun 2021 15:13:39 GMT (592kb,D)
[v2] Thu, 1 Jul 2021 13:41:16 GMT (592kb,D)
[v3] Wed, 24 Nov 2021 20:09:15 GMT (59kb)

Link back to: arXiv, form interface, contact.