Buchumschlag

Beyond one-hot encoding: lower dimensional target embedding

Target encoding plays a central role when learning Convolutional Neural Networks. In this realm, one-hot encoding is the most prevalent strategy due to its simplicity. However, this so widespread encoding schema assumes a flat label space, thus ignoring rich relationships existing among labels that...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Rodriguez, Pau (VerfasserIn) , Bautista, Miguel (VerfasserIn) , Gonzàlez, Jordi (VerfasserIn) , Escalera, Sergio (VerfasserIn)
Dokumenttyp: Article (Journal)
Sprache:Englisch
Veröffentlicht: 11 May 2018
In: Image and vision computing
Year: 2018, Jahrgang: 75, Pages: 21-31
DOI:10.1016/j.imavis.2018.04.004
Online-Zugang:Verlag, lizenzpflichtig, Volltext: https://doi.org/10.1016/j.imavis.2018.04.004
Verlag, lizenzpflichtig, Volltext: http://www.sciencedirect.com/science/article/pii/S0262885618300623
Volltext
Verfasserangaben:Pau Rodríguez, Miguel A. Bautista, Jordi Gonzàlez, Sergio Escalera

MARC

LEADER 00000caa a2200000 c 4500
001 1695324978
003 DE-627
005 20220818045917.0
007 cr uuu---uuuuu
008 200421s2018 xx |||||o 00| ||eng c
024 7 |a 10.1016/j.imavis.2018.04.004  |2 doi 
035 |a (DE-627)1695324978 
035 |a (DE-599)KXP1695324978 
035 |a (OCoLC)1341316057 
040 |a DE-627  |b ger  |c DE-627  |e rda 
041 |a eng 
084 |a 28  |2 sdnb 
100 1 |a Rodriguez, Pau  |e VerfasserIn  |0 (DE-588)1208730339  |0 (DE-627)1695350812  |4 aut 
245 1 0 |a Beyond one-hot encoding  |b lower dimensional target embedding  |c Pau Rodríguez, Miguel A. Bautista, Jordi Gonzàlez, Sergio Escalera 
264 1 |c 11 May 2018 
300 |a 11 
336 |a Text  |b txt  |2 rdacontent 
337 |a Computermedien  |b c  |2 rdamedia 
338 |a Online-Ressource  |b cr  |2 rdacarrier 
500 |a Gesehen am 21.04.2020 
520 |a Target encoding plays a central role when learning Convolutional Neural Networks. In this realm, one-hot encoding is the most prevalent strategy due to its simplicity. However, this so widespread encoding schema assumes a flat label space, thus ignoring rich relationships existing among labels that can be exploited during training. In large-scale datasets, data does not span the full label space, but instead lies in a low-dimensional output manifold. Following this observation, we embed the targets into a low-dimensional space, drastically improving convergence speed while preserving accuracy. Our contribution is two fold: (i) We show that random projections of the label space are a valid tool to find such lower dimensional embeddings, boosting dramatically convergence rates at zero computational cost; and (ii) we propose a normalized eigenrepresentation of the class manifold that encodes the targets with minimal information loss, improving the accuracy of random projections encoding while enjoying the same convergence rates. Experiments on CIFAR-100, CUB200-2011, Imagenet, and MIT Places demonstrate that the proposed approach drastically improves convergence speed while reaching very competitive accuracy rates. 
650 4 |a Computer vision 
650 4 |a Deep learning 
650 4 |a Error correcting output codes 
650 4 |a Output embeddings 
700 1 |a Bautista, Miguel  |e VerfasserIn  |0 (DE-588)115223787X  |0 (DE-627)1013721810  |0 (DE-576)49975705X  |4 aut 
700 1 |a Gonzàlez, Jordi  |e VerfasserIn  |4 aut 
700 1 |a Escalera, Sergio  |e VerfasserIn  |4 aut 
773 0 8 |i Enthalten in  |t Image and vision computing  |d Amsterdam [u.a.] : Elsevier Science, 1983  |g 75(2018), Seite 21-31  |h Online-Ressource  |w (DE-627)270938087  |w (DE-600)1478755-6  |w (DE-576)078316820  |7 nnas  |a Beyond one-hot encoding lower dimensional target embedding 
773 1 8 |g volume:75  |g year:2018  |g pages:21-31  |g extent:11  |a Beyond one-hot encoding lower dimensional target embedding 
856 4 0 |u https://doi.org/10.1016/j.imavis.2018.04.004  |x Verlag  |x Resolving-System  |z lizenzpflichtig  |3 Volltext 
856 4 0 |u http://www.sciencedirect.com/science/article/pii/S0262885618300623  |x Verlag  |z lizenzpflichtig  |3 Volltext 
951 |a AR 
992 |a 20200421 
993 |a Article 
994 |a 2018 
998 |g 115223787X  |a Bautista, Miguel  |m 115223787X:Bautista, Miguel  |d 700000  |d 708070  |e 700000PB115223787X  |e 708070PB115223787X  |k 0/700000/  |k 1/700000/708070/  |p 2 
999 |a KXP-PPN1695324978  |e 3627845811 
BIB |a Y 
SER |a journal 
JSO |a {"physDesc":[{"extent":"11 S."}],"relHost":[{"language":["eng"],"recId":"270938087","note":["Gesehen am 03.02.15"],"disp":"Beyond one-hot encoding lower dimensional target embeddingImage and vision computing","type":{"bibl":"periodical","media":"Online-Ressource"},"part":{"volume":"75","text":"75(2018), Seite 21-31","extent":"11","year":"2018","pages":"21-31"},"pubHistory":["1.1983 - 32.2014; Vol. 33.2015 -"],"title":[{"title_sort":"Image and vision computing","title":"Image and vision computing"}],"physDesc":[{"extent":"Online-Ressource"}],"id":{"eki":["270938087"],"zdb":["1478755-6"]},"origin":[{"publisherPlace":"Amsterdam [u.a.]","publisher":"Elsevier Science","dateIssuedKey":"1983","dateIssuedDisp":"1983-"}]}],"name":{"displayForm":["Pau Rodríguez, Miguel A. Bautista, Jordi Gonzàlez, Sergio Escalera"]},"origin":[{"dateIssuedDisp":"11 May 2018","dateIssuedKey":"2018"}],"id":{"doi":["10.1016/j.imavis.2018.04.004"],"eki":["1695324978"]},"note":["Gesehen am 21.04.2020"],"type":{"media":"Online-Ressource","bibl":"article-journal"},"language":["eng"],"recId":"1695324978","person":[{"role":"aut","display":"Rodriguez, Pau","roleDisplay":"VerfasserIn","given":"Pau","family":"Rodriguez"},{"roleDisplay":"VerfasserIn","display":"Bautista, Miguel","role":"aut","family":"Bautista","given":"Miguel"},{"given":"Jordi","family":"Gonzàlez","role":"aut","display":"Gonzàlez, Jordi","roleDisplay":"VerfasserIn"},{"family":"Escalera","given":"Sergio","display":"Escalera, Sergio","roleDisplay":"VerfasserIn","role":"aut"}],"title":[{"title":"Beyond one-hot encoding","subtitle":"lower dimensional target embedding","title_sort":"Beyond one-hot encoding"}]} 
SRT |a RODRIGUEZPBEYONDONEH1120 

Ähnliche Einträge