國立虎尾科技大學 |

Data Systems for Explainable AI and Incorporating AI Infrastructure Into Data Systems /

紀錄類型:	書目-語言資料,印刷品 : Monograph/item
正題名/作者:	Data Systems for Explainable AI and Incorporating AI Infrastructure Into Data Systems // Dong He.
作者:	He, Dong,
面頁冊數:	1 electronic resource (193 pages)
附註:	Source: Dissertations Abstracts International, Volume: 86-03, Section: B.
Contained By:	Dissertations Abstracts International86-03B.
標題:	Information technology. -
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31556793
ISBN:	9798384094005

Data Systems for Explainable AI and Incorporating AI Infrastructure Into Data Systems /
He, Dong,

Data Systems for Explainable AI and Incorporating AI Infrastructure Into Data Systems /Dong He. - 1 electronic resource (193 pages)

Source: Dissertations Abstracts International, Volume: 86-03, Section: B.

Artificial Intelligence (AI) has become a cornerstone of modern computing, powering a wide range of applications in fields from face recognition and machine translation to medical diagnosis and autonomous driving. This transformation is advancing data-driven AI models that not only learn from vast amounts of data but can generate substantial data artifacts. These changes introduce significant data management challenges, particularly as AI models grow in architectural complexity and data intensity, because traditional data management systems were not designed to handle AI workloads.Concurrently, the tremendous computational and memory demands of AI workloads have driven the rapid development of specialized AI hardware and software infrastructure. This evolution has created systems that not only accelerate AI workloads but offer new opportunities to improve the performance of relational query processing in data management systems. Despite these advances, the diversity in hardware characteristics and programming abstractions, coupled with the lack of native support in traditional data management systems, presents significant challenges for data management system builders to fully leverage the exciting potential of such AI infrastructure.This dissertation aims to bridge the worlds of AI and data management by introducing new data systems that efficiently support explainable AI, a subset of especially data-intensive AI workloads, and leveraging AI infrastructure to accelerate relational query processing in data management systems.First, we introduce two data systems for efficient explainable AI: DeepEverest and MaskSearch. Each system supports a different type of AI model explanation task. DeepEverest accelerates neural network explanation queries that return input examples with certain neuron activation patterns; these queries help practitioners understand the functionality of groups of neurons in a neural network by tying that functionality to the input examples. MaskSearch enables efficient querying over databases of image masks generated by AI models (e.g., segmentation masks, saliency maps, etc.), supporting the retrieval of masks with particular characteristics that are crucial for applications such as identifying spurious correlations, detecting adversarial examples, and monitoring model errors.Second, we introduce the Tensor Query Processor (TQP), the industry's first query processor that compiles SQL queries into tensor programs (i.e., PyTorch programs) and executes them on any hardware backend supported by the tensor runtime, including CPUs, GPUs, and TPUs. TQP demonstrates the potential of using AI infrastructure to accelerate relational query processing in data management systems by supporting the full TPC-H benchmark and outperforming state-of-the-art systems. Further, it bridges the gap between AI workloads and relational queries by providing a unified intermediate representation for efficient execution when both types of workloads are present in the same system. While much work remains to be done, this dissertation contributes an important step towards improving data management systems in the novel era of AI.

English

ISBN: 9798384094005Subjects--Topical Terms:

559429
Information technology.
Subjects--Index Terms:

Data management

Data Systems for Explainable AI and Incorporating AI Infrastructure Into Data Systems /
LDR:04672nam a22004453i 4500 001 1157861
005 20250603111430.5
006 m o d
007 cr|nu||||||||
008 250804s2024 miu||||||m |||||||eng d
020 $a 9798384094005
035 $a (MiAaPQD)AAI31556793
035 $a AAI31556793
040 $a MiAaPQD $b eng $c MiAaPQD $e rda
100 1 $a He, Dong, $e author. $3 1484144
245 1 0 $a Data Systems for Explainable AI and Incorporating AI Infrastructure Into Data Systems / $c Dong He.
264 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2024
300 $a 1 electronic resource (193 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: 86-03, Section: B.
500 $a Advisors: Balazinska, Magdalena Committee members: Suciu, Dan; Krishna, Ranjay.
502 $b Ph.D. $c University of Washington $d 2024.
520 $a Artificial Intelligence (AI) has become a cornerstone of modern computing, powering a wide range of applications in fields from face recognition and machine translation to medical diagnosis and autonomous driving. This transformation is advancing data-driven AI models that not only learn from vast amounts of data but can generate substantial data artifacts. These changes introduce significant data management challenges, particularly as AI models grow in architectural complexity and data intensity, because traditional data management systems were not designed to handle AI workloads.Concurrently, the tremendous computational and memory demands of AI workloads have driven the rapid development of specialized AI hardware and software infrastructure. This evolution has created systems that not only accelerate AI workloads but offer new opportunities to improve the performance of relational query processing in data management systems. Despite these advances, the diversity in hardware characteristics and programming abstractions, coupled with the lack of native support in traditional data management systems, presents significant challenges for data management system builders to fully leverage the exciting potential of such AI infrastructure.This dissertation aims to bridge the worlds of AI and data management by introducing new data systems that efficiently support explainable AI, a subset of especially data-intensive AI workloads, and leveraging AI infrastructure to accelerate relational query processing in data management systems.First, we introduce two data systems for efficient explainable AI: DeepEverest and MaskSearch. Each system supports a different type of AI model explanation task. DeepEverest accelerates neural network explanation queries that return input examples with certain neuron activation patterns; these queries help practitioners understand the functionality of groups of neurons in a neural network by tying that functionality to the input examples. MaskSearch enables efficient querying over databases of image masks generated by AI models (e.g., segmentation masks, saliency maps, etc.), supporting the retrieval of masks with particular characteristics that are crucial for applications such as identifying spurious correlations, detecting adversarial examples, and monitoring model errors.Second, we introduce the Tensor Query Processor (TQP), the industry's first query processor that compiles SQL queries into tensor programs (i.e., PyTorch programs) and executes them on any hardware backend supported by the tensor runtime, including CPUs, GPUs, and TPUs. TQP demonstrates the potential of using AI infrastructure to accelerate relational query processing in data management systems by supporting the full TPC-H benchmark and outperforming state-of-the-art systems. Further, it bridges the gap between AI workloads and relational queries by providing a unified intermediate representation for efficient execution when both types of workloads are present in the same system. While much work remains to be done, this dissertation contributes an important step towards improving data management systems in the novel era of AI.
546 $a English
590 $a School code: 0250
650 4 $a Information technology. $3 559429
650 4 $a Engineering. $3 561152
650 4 $a Computer science. $3 573171
653 $a Data management
653 $a AI models
653 $a MaskSearch
653 $a Tensor Query Processor
653 $a Programming abstractions
690 $a 0984
690 $a 0489
690 $a 0800
690 $a 0537
710 2 $a University of Washington. $b Computer Science and Engineering. $3 1182238
720 1 $a Balazinska, Magdalena $e degree supervisor.
773 0 $t Dissertations Abstracts International $g 86-03B.
790 $a 0250
791 $a Ph.D.
792 $a 2024
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31556793