國立虎尾科技大學 |

Topology-Driven Learning for Images: Applications and Acceleration /

紀錄類型:	書目-語言資料,印刷品 : Monograph/item
正題名/作者:	Topology-Driven Learning for Images: Applications and Acceleration // Fan Wang.
作者:	Wang, Fan,
面頁冊數:	1 electronic resource (141 pages)
附註:	Source: Dissertations Abstracts International, Volume: 86-05, Section: A.
Contained By:	Dissertations Abstracts International86-05A.
標題:	Computer science. -
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31481887
ISBN:	9798342728478

Topology-Driven Learning for Images: Applications and Acceleration /
Wang, Fan,

Topology-Driven Learning for Images: Applications and Acceleration /Fan Wang. - 1 electronic resource (141 pages)

Source: Dissertations Abstracts International, Volume: 86-05, Section: A.

The integration of topological methods into deep learning, particularly through persistent homology, has seen significant growth recently. Persistent homology, a key concept in Topological Data Analysis (TDA), offers insight into the topological characteristics of data by tracing their evolution across scales. This thesis explores the application of persistent homology to deep learning challenges, specifically focusing on ensuring the topological integrity of generated images. We enhance the robustness of a generative adversarial network by introducing a penalty based on the disparity in topological properties between the generated images and the training distribution. This approach ensures that the generated images resemble the training data not only visually but also in their topological characteristics.As the second application, We extract persistent homology from breast DEC-MRI volumes as an approximation of fibroglandular tissues and use it to explicitly direct the attention of a 3D network to a smaller set of voxels with high biological relevance. This targeted approach allows the network to focus on areas that are most indicative of underlying pathologies.Despite the practical benefits, challenges remain in melding persistent homology with deep learning due to computational demands. Addressing this, the research focuses on massively parallel GPU algorithms to expedite persistent homology calculations, aiming to bridge this gap in deep learning frameworks. The shift from CPU to GPU computing is driven by the diminishing returns of Moore's Law, prompting a transition to accelerated computing with GPUs leading the performance enhancement.This research delves into persistent homology computation for image data-essential in fields ranging from medical imaging to physical simulations, and crucial in deep learning as convolutional network outputs. This thesis introduces a streaming GPU algorithm for computing persistent homology in 2D and 3D digital images, addressing a notable bottleneck in deep learning training times. By constructing a cubical complex and utilizing discrete Morse theory for gradient vector field development, the proposed algorithm applies innovative massively parallel algorithms for topological sorting and path parity to formulate Morse boundaries. It assembles partial boundary relations from discrete data segments into a unified global boundary matrix for reduction, employing specialized measures to manage Morse matchings at chunk and GPU borders. This ensures the fidelity of persistent homology within a streaming context. The resulting algorithm exhibits exceptional speed improvements in preprocessing and persistent homology computations, surpassing existing state-of-the-art methods.This work aims to contribute substantially improved algorithms to the TDA community, significantly reducing computation times and fostering further applications of topological loss in deep learning, thereby removing computational barriers and advancing the field.

English

ISBN: 9798342728478Subjects--Topical Terms:

573171
Computer science.
Subjects--Index Terms:

Generative adversarial network

Topology-Driven Learning for Images: Applications and Acceleration /
LDR:04641nam a22004693i 4500 001 1157841
005 20250603111426.5
006 m o d
007 cr|nu||||||||
008 250804s2024 miu||||||m |||||||eng d
020 $a 9798342728478
035 $a (MiAaPQD)AAI31481887
035 $a AAI31481887
040 $a MiAaPQD $b eng $c MiAaPQD $e rda
100 1 $a Wang, Fan, $e author. $3 1484119
245 1 0 $a Topology-Driven Learning for Images: Applications and Acceleration / $c Fan Wang.
264 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2024
300 $a 1 electronic resource (141 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: 86-05, Section: A.
500 $a Advisors: Chen, Chao; Samaras, Dimitris Committee members: Chowdhury, Rezaul A.; Ling, Haibin; Wagner, Hubert.
502 $b Ph.D. $c State University of New York at Stony Brook $d 2024.
520 $a The integration of topological methods into deep learning, particularly through persistent homology, has seen significant growth recently. Persistent homology, a key concept in Topological Data Analysis (TDA), offers insight into the topological characteristics of data by tracing their evolution across scales. This thesis explores the application of persistent homology to deep learning challenges, specifically focusing on ensuring the topological integrity of generated images. We enhance the robustness of a generative adversarial network by introducing a penalty based on the disparity in topological properties between the generated images and the training distribution. This approach ensures that the generated images resemble the training data not only visually but also in their topological characteristics.As the second application, We extract persistent homology from breast DEC-MRI volumes as an approximation of fibroglandular tissues and use it to explicitly direct the attention of a 3D network to a smaller set of voxels with high biological relevance. This targeted approach allows the network to focus on areas that are most indicative of underlying pathologies.Despite the practical benefits, challenges remain in melding persistent homology with deep learning due to computational demands. Addressing this, the research focuses on massively parallel GPU algorithms to expedite persistent homology calculations, aiming to bridge this gap in deep learning frameworks. The shift from CPU to GPU computing is driven by the diminishing returns of Moore's Law, prompting a transition to accelerated computing with GPUs leading the performance enhancement.This research delves into persistent homology computation for image data-essential in fields ranging from medical imaging to physical simulations, and crucial in deep learning as convolutional network outputs. This thesis introduces a streaming GPU algorithm for computing persistent homology in 2D and 3D digital images, addressing a notable bottleneck in deep learning training times. By constructing a cubical complex and utilizing discrete Morse theory for gradient vector field development, the proposed algorithm applies innovative massively parallel algorithms for topological sorting and path parity to formulate Morse boundaries. It assembles partial boundary relations from discrete data segments into a unified global boundary matrix for reduction, employing specialized measures to manage Morse matchings at chunk and GPU borders. This ensures the fidelity of persistent homology within a streaming context. The resulting algorithm exhibits exceptional speed improvements in preprocessing and persistent homology computations, surpassing existing state-of-the-art methods.This work aims to contribute substantially improved algorithms to the TDA community, significantly reducing computation times and fostering further applications of topological loss in deep learning, thereby removing computational barriers and advancing the field.
546 $a English
590 $a School code: 0771
650 4 $a Computer science. $3 573171
650 4 $a Linguistics. $3 557829
650 4 $a Information technology. $3 559429
653 $a Generative adversarial network
653 $a GPU computation
653 $a Parallel computation
653 $a Persistent homology
653 $a Topological Data Analysis
653 $a Topology
690 $a 0984
690 $a 0489
690 $a 0800
690 $a 0290
710 2 $a State University of New York at Stony Brook. $b Computer Science. $3 1180378
720 1 $a Chen, Chao $e degree supervisor.
720 1 $a Samaras, Dimitris $e degree supervisor.
773 0 $t Dissertations Abstracts International $g 86-05A.
790 $a 0771
791 $a Ph.D.
792 $a 2024
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31481887