國立虎尾科技大學 |

Learning Activation Functions in Deep Neural Networks.

紀錄類型:	書目-語言資料,手稿 : Monograph/item
正題名/作者:	Learning Activation Functions in Deep Neural Networks./
作者:	Farhadi, Farnoush.
面頁冊數:	1 online resource (172 pages)
附註:	Source: Dissertation Abstracts International, Volume: 76-11C.
Contained By:	Dissertation Abstracts International76-11C.
標題:	Industrial engineering. -
電子資源:	click for full text (PQDT)

Learning Activation Functions in Deep Neural Networks.
Farhadi, Farnoush.

Learning Activation Functions in Deep Neural Networks. - 1 online resource (172 pages)

Source: Dissertation Abstracts International, Volume: 76-11C.

Thesis (M.A.Sc.)--Ecole Polytechnique, Montreal (Canada), 2017.

Includes bibliographical references

Recently, user intention prediction based on behavioral data has become more attractive in digital marketing and online advertisement. The prediction is often performed by analyzing the logged information which includes details on how each user visited the commercial website. In this work, we explore the machine learning methods for classification of visitors to an e-commerce website based on URLs they visited. We aim at studying on how effectively the state-of-the-art deep neural networks could predict the purpose of visitors in a session using only URL sequences. Typical deep neural networks employ a fixed nonlinear activation function for each hidden neuron. In this thesis, two adaptive activation functions for individual hidden units are proposed such that the deep network learns these activation functions in training. Methods and algorithms for developing these adaptive activation functions are discussed. Furthermore, performance of the proposed activation functions in deep networks compared to state-of-the-art models are also evaluated on a real-world URL dataset and two well-known benchmarks, MNIST and Movie Review benchmarks as well.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2018

Mode of access: World Wide Web

Subjects--Topical Terms:

679492
Industrial engineering.
Index Terms--Genre/Form:

554714
Electronic books.

Learning Activation Functions in Deep Neural Networks.
LDR:04583ntm a2200337Ki 4500 001 916939
005 20180928111503.5
006 m o u
007 cr mn||||a|a||
008 190606s2017 xx obm 000 0 eng d
035 $a (MiAaPQ)AAI10957109
035 $a (MiAaPQ)EPMontreal2945
035 $a AAI10957109
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Farhadi, Farnoush. $3 1190815
245 1 0 $a Learning Activation Functions in Deep Neural Networks.
264 0 $c 2017
300 $a 1 online resource (172 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertation Abstracts International, Volume: 76-11C.
500 $a Advisers: Andrea Lodi; Vahid Partovi Nia.
502 $a Thesis (M.A.Sc.)--Ecole Polytechnique, Montreal (Canada), 2017.
504 $a Includes bibliographical references
520 $a Recently, user intention prediction based on behavioral data has become more attractive in digital marketing and online advertisement. The prediction is often performed by analyzing the logged information which includes details on how each user visited the commercial website. In this work, we explore the machine learning methods for classification of visitors to an e-commerce website based on URLs they visited. We aim at studying on how effectively the state-of-the-art deep neural networks could predict the purpose of visitors in a session using only URL sequences. Typical deep neural networks employ a fixed nonlinear activation function for each hidden neuron. In this thesis, two adaptive activation functions for individual hidden units are proposed such that the deep network learns these activation functions in training. Methods and algorithms for developing these adaptive activation functions are discussed. Furthermore, performance of the proposed activation functions in deep networks compared to state-of-the-art models are also evaluated on a real-world URL dataset and two well-known benchmarks, MNIST and Movie Review benchmarks as well.
520 $a Recemment, la prediction de l'intention des utilisateurs basee sur les donnees comportementales est devenue plus attrayante dans les domaines de marketing numerique et publicite en ligne. La prediction est souvent effectuee par analyse de l'information du comportement des visiteurs qui comprend des details sur la facon dont chaque utilisateur a visite le site Web commercial. Dans ce travail, nous explorons les methodes statistiques d'apprentissage automatique pour classifier les visiteurs d'un site Web de commerce electronique d'apres des URL qu'ils ont parcourues. Plus precisement, on vise a etudier comment les reseaux de neurones profonds de pointe pourraient efficacement predire le but des visiteurs dans une session de visite, en utilisant uniquement des sequences d'URL. Dans un reseau contenant des couches limitees, le choix de la fonction d'activation a un effet sur l'apprentissage de la representation et les performances du reseau. Typiquement, les modeles de reseaux de neurones sont appris en utilisant des fonctions d'activation specifiques comme la fonction sigmoide symetrique standard ou relu lineaire. Dans ce memoire, on vise a proposer une fonction sigmoide asymetrique parametree dont la parametre pourrait etre controle et ajuste dans chacun des neurons cachees avec d'autres parametres de reseau profond en formation. De plus, nous visons a proposer une variante non-lineaire de la fonction relu pour les reseaux profonds. Comme le sigmoide adaptatif, notre objectif est de regler les parametres du relu adaptatif pour chaque unite individuellement, dans l'etape de formation. Des methodes et des algorithmes pour developper ces fonctions d'activation adaptatives sont discutes. En outre, une petite variante de MLP (Multi Layer Perceptron) et un modele CNN (Convolutional Neural Network) appliquant nos fonctions d'activation proposees sont utilises pour predire l'intention des utilisateurs selon les donnees d'URL. Quatre jeux de donnees differents ont ete choisis, appele les donnees simulees, les donnees MNIST, les donnees de revue de film, et les donnees d'URL pour demontrer l'effet de selectionner differentes fonctions d'activation sur les modeles MLP et CNN proposes.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2018
538 $a Mode of access: World Wide Web
650 4 $a Industrial engineering. $3 679492
650 4 $a Artificial intelligence. $3 559380
655 7 $a Electronic books. $2 local $3 554714
690 $a 0546
690 $a 0800
710 2 $a ProQuest Information and Learning Co. $3 1178819
710 2 $a Ecole Polytechnique, Montreal (Canada). $3 845478
773 0 $t Dissertation Abstracts International $g 76-11C.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=10957109 $z click for full text (PQDT)