Derivations of the Enhanced Gradient for the Boltzmann Machine

No Thumbnail Available

URL

Journal Title

Journal ISSN

Volume Title

School of Science | D4 Julkaistu kehittämis- tai tutkimusraportti tai -selvitys

Date

2011

Major/Subject

Mcode

Degree programme

Language

en

Pages

9

Series

Aalto University publication series SCIENCE+TECHNOLOGY, 20/2011

Abstract

This technical report is extends the conference paper [1] and the abstract [2] with detailed derivations and proofs. First we recap notation that we use on the Boltzmann machine and its learning. Then we define transformations for the machine where some of its bits are flipped for all samples, and show the equivalence of the transformed model to the original one. Then we show that traditional update rules are not invariant to the transformations, propose a new update rule called theenhanced gradient, and finally show its invariance to the transformations.

Description

Keywords

enhanced gradient, boltzmann machine, sample index

Other note

Citation