Stochastic Shadow-Cutting Machine

No Thumbnail Available

Access rights

openAccess
acceptedVersion

URL

Journal Title

Journal ISSN

Volume Title

A4 Artikkeli konferenssijulkaisussa

Date

2024-01-01

Major/Subject

Mcode

Degree programme

Language

en

Pages

4

Series

2023 31st Telecommunications Forum, TELFOR 2023 - Proceedings

Abstract

Recently, a new concept called shadow-cuts has recently been proposed for a fully-connected graph whose edge matrix is Hermitian with arbitrary complex numbers. Each neuron is associated with a phase and the sum of shadow cuts is defined as the sum of inter-cluster phased edges. However, the shadow-cut machine is 100% deterministic and therefore its modeling capacity is relatively limited. In this brief, we (i) extend it to stochastic domain which yields the so-called 'Stochastic Shadow-Cutting Machine' (SSCM), and (ii) show that choosing the energy function of the SSCM as the sum of shadow-cuts yields similar phenomena as in those from the statistical mechanics like Ising model, xy-model, pott model, Stochastic Hopfield Networks, etc., Thus, the proposed SSCM provides a general framework to examine various phenomena like the phase changes of the SSCM as the temperature increases. Because the SSCM in low temperatures behaves as an Associative Memory system (i.e., 'ferro-magnet'), it is possible to examine the critical temperatures when the SSCM cannot 'recover/remember' the patterns any more (i.e. 'anti-ferromagnet'), which we define as 'phase change' of the SSCM.

Description

Publisher Copyright: © 2023 IEEE.

Keywords

associative memory systems, Graphs with complex-valued edges, inter-cluster phased edges, Ising model, statistical mechanics, Stochastic Shadow-Cutting Machine

Other note

Citation

Uykan, Z & Jantti, R 2024, Stochastic Shadow-Cutting Machine . in 2023 31st Telecommunications Forum, TELFOR 2023 - Proceedings . 2023 31st Telecommunications Forum, TELFOR 2023 - Proceedings, IEEE, Telecommunications Forum, Belgrade, Serbia, 21/11/2023 . https://doi.org/10.1109/TELFOR59449.2023.10372707