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: Analysis and Design of Quadratic Neural Networks for Regression, Classification, and Lyapunov Control of Dynamical Systems

Abstract: This paper addresses the analysis and design of quadratic neural networks, which have been recently introduced in the literature, and their applications to regression, classification, system identification and control of dynamical systems. These networks offer several advantages, the most important of which are the fact that the architecture is a by-product of the design and is not determined a-priori, their training can be done by solving a convex optimization problem so that the global optimum of the weights is achieved, and the input-output mapping can be expressed analytically by a quadratic form. It also appears from several examples that these networks work extremely well using only a small fraction of the training data. The results in the paper cast regression, classification, system identification, stability and control design as convex optimization problems, which can be solved efficiently with polynomial-time algorithms to a global optimum. Several examples will show the effectiveness of quadratic neural networks in applications.
Comments: submitted in May 2022
Subjects: Machine Learning (cs.LG); Optimization and Control (math.OC)
Cite as: arXiv:2207.13120 [cs.LG]
  (or arXiv:2207.13120v1 [cs.LG] for this version)

Submission history

From: Luis Rodrigues [view email]
[v1] Tue, 26 Jul 2022 18:10:05 GMT (219kb,D)

Link back to: arXiv, form interface, contact.