A new approach for cells in multidimensional recurrent neural networks

Gundram Leifert; Tobias Strauß; Roger Labahn; Welf Wustlich

doi:10.18453/rosdok_id00005819

zum übergeordneten Dokument

Gundram Leifert Tobias Strauß Roger Labahn Welf Wustlich

A new approach for cells in multidimensional recurrent neural networks

University of Rostock, Institute of Mathematics, 2012

https://doi.org/10.18453/rosdok_id00005819

Abstract: A recent approach for offline handwriting recognition is to use multidimensional recurrent neural networks (MDRNN) with connectionist temporal classification which has shown to yield very good results on several datasets. MDRNNs contain special units – multidimensional Long Short-Term Memory (MDLSTM) cells. These cells suffer from instability especially for higher dimensionality. We analyze the reasons for this effect and introduce several cells with better stability. We present a method to design stable multidimensional cells using the theory of linear shift invariant systems. The new cells are compared to MDLSTMs on the Arabic and French ICDAR datasets, where they yield better results.

Workingpaper Freier Zugang

Titel:

A new approach for cells in multidimensional recurrent neural networks

Schriftenreihe:

Preprints aus dem Institut für Mathematik, 12/4

Beteiligte Personen:

Gundram Leifert[VerfasserIn]
Tobias Strauß[VerfasserIn]
	0000-0002-5344-4188
	1136294554
Roger Labahn[VerfasserIn]
	0000-0003-1901-9644
	170999645
Welf Wustlich[VerfasserIn]

Zusammenfassung:

A recent approach for offline handwriting recognition is to use multidimensional recurrent neural networks (MDRNN) with connectionist temporal classification which has shown to yield very good results on several datasets. MDRNNs contain special units – multidimensional Long Short-Term Memory (MDLSTM) cells. These cells suffer from instability especially for higher dimensionality. We analyze the reasons for this effect and introduce several cells with better stability. We present a method to design stable multidimensional cells using the theory of linear shift invariant systems. The new cells are compared to MDLSTMs on the Arabic and French ICDAR datasets, where they yield better results. [Englisch]

Dokumenttyp:

Workingpaper

Einrichtung:

Mathematisch-Naturwissenschaftliche Fakultät

Sprache:

Englisch

Sachgruppe der DNB:

510 Mathematik

Umfang:

1 Online-Ressource (22 Seiten)

Veröffentlichung /
Entstehung:

Rostock: University of Rostock, Institute of Mathematics