國立虎尾科技大學 |

Information Flow in a Deep Learning Classification Network.

紀錄類型:	書目-語言資料,手稿 : Monograph/item
正題名/作者:	Information Flow in a Deep Learning Classification Network./
作者:	Banerjee, Debanjali.
面頁冊數:	1 online resource (77 pages)
附註:	Source: Dissertations Abstracts International, Volume: 85-12, Section: B.
Contained By:	Dissertations Abstracts International85-12B.
標題:	Computer science. -
電子資源:	click for full text (PQDT)
ISBN:	9798383179512

Information Flow in a Deep Learning Classification Network.
Banerjee, Debanjali.

Information Flow in a Deep Learning Classification Network. - 1 online resource (77 pages)

Source: Dissertations Abstracts International, Volume: 85-12, Section: B.

Thesis (Ph.D.)--University of Louisiana at Lafayette, 2024.

Includes bibliographical references

Deep learning methods have excellent accuracy achievements in image classification but largely remains a black box method. Image classification is the core of many machine vision tasks, including object detection. Object detection tasks rely on understanding how classifiers make their decisions. By training a deep learning network to differentiate between two classes and then computing SHapley Additive ExPlanations (SHAP) values for its feature layers, I analyze how the network 'sees' the input images. SHAP values, derived from the Shapley value in game theory, reveal how much each feature contributes to the final classification decision. Importantly, they adhere to principles of local accuracy, missingness, and consistency. My experiments demonstrate that the feature extraction layers within the network exhibit distinct response patterns that correspond to the shape of the target objects, at levels earlier than what was previously known. Conventional wisdom in deep learning suggests that early layers of convolutional neural networks (CNNs) extract general features, while middle layers develop class-specific representations. However, we lack strong empirical evidence to support precisely how this information transformation occurs. I use the two-class classification task and measure the self- and cross-entropies at each feature node of the deep learning network. From these entropy values I derive the Jensen-Shannon Divergence at each feature node. The Interquartile Range of these divergence values at a layer then serves as an indicator of the divergence of the representations of the two classes. My experiments reveal that class representations become increasingly distinct as we progress through a network's layers. This supports the idea that deeper layers learn more complex and discriminative features, crucial for accurate classification.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2024

Mode of access: World Wide Web

ISBN: 9798383179512Subjects--Topical Terms:

573171
Computer science.
Subjects--Index Terms:

Convolutional neural networksIndex Terms--Genre/Form:

554714
Electronic books.

Information Flow in a Deep Learning Classification Network.
LDR:03274ntm a22003977 4500 001 1150196
005 20241022111612.5
006 m o d
007 cr bn ---uuuuu
008 250605s2024 xx obm 000 0 eng d
020 $a 9798383179512
035 $a (MiAaPQ)AAI31294982
035 $a AAI31294982
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Banerjee, Debanjali. $3 1476637
245 1 0 $a Information Flow in a Deep Learning Classification Network.
264 0 $c 2024
300 $a 1 online resource (77 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: Dissertations Abstracts International, Volume: 85-12, Section: B.
500 $a Advisor: Chu, Chee-Hung Henry.
502 $a Thesis (Ph.D.)--University of Louisiana at Lafayette, 2024.
504 $a Includes bibliographical references
520 $a Deep learning methods have excellent accuracy achievements in image classification but largely remains a black box method. Image classification is the core of many machine vision tasks, including object detection. Object detection tasks rely on understanding how classifiers make their decisions. By training a deep learning network to differentiate between two classes and then computing SHapley Additive ExPlanations (SHAP) values for its feature layers, I analyze how the network 'sees' the input images. SHAP values, derived from the Shapley value in game theory, reveal how much each feature contributes to the final classification decision. Importantly, they adhere to principles of local accuracy, missingness, and consistency. My experiments demonstrate that the feature extraction layers within the network exhibit distinct response patterns that correspond to the shape of the target objects, at levels earlier than what was previously known. Conventional wisdom in deep learning suggests that early layers of convolutional neural networks (CNNs) extract general features, while middle layers develop class-specific representations. However, we lack strong empirical evidence to support precisely how this information transformation occurs. I use the two-class classification task and measure the self- and cross-entropies at each feature node of the deep learning network. From these entropy values I derive the Jensen-Shannon Divergence at each feature node. The Interquartile Range of these divergence values at a layer then serves as an indicator of the divergence of the representations of the two classes. My experiments reveal that class representations become increasingly distinct as we progress through a network's layers. This supports the idea that deeper layers learn more complex and discriminative features, crucial for accurate classification.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2024
538 $a Mode of access: World Wide Web
650 4 $a Computer science. $3 573171
650 4 $a Information technology. $3 559429
653 $a Convolutional neural networks
653 $a Deep learning methods
653 $a Image classification
653 $a Accuracy achievements
653 $a Black box method
655 7 $a Electronic books. $2 local $3 554714
690 $a 0984
690 $a 0800
690 $a 0489
710 2 $a ProQuest Information and Learning Co. $3 1178819
710 2 $a University of Louisiana at Lafayette. $b Computer Science. $3 1195497
773 0 $t Dissertations Abstracts International $g 85-12B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31294982 $z click for full text (PQDT)