國立虎尾科技大學 |

Accessible Artificial Intelligence.

紀錄類型:	書目-語言資料,手稿 : Monograph/item
正題名/作者:	Accessible Artificial Intelligence./
作者:	Bingham, Joseph.
面頁冊數:	1 online resource (151 pages)
附註:	Source: Dissertations Abstracts International, Volume: 85-05, Section: B.
Contained By:	Dissertations Abstracts International85-05B.
標題:	Computer engineering. -
電子資源:	click for full text (PQDT)
ISBN:	9798380845793

Accessible Artificial Intelligence.
Bingham, Joseph.

Accessible Artificial Intelligence. - 1 online resource (151 pages)

Source: Dissertations Abstracts International, Volume: 85-05, Section: B.

Thesis (Ph.D.)--Rutgers The State University of New Jersey, School of Graduate Studies, 2023.

Includes bibliographical references

As science reaches forward with novel and greater developments, the gap between what the average individual knows that the knowledge required to understand the advanced technology that they interact with. This gap is felt in many fields, and has recently lead to a sociological phenomenon known as truth decay. Jennifer Kavanagh of RAND defines truth decay as the diminishing role that data and expert analysis effect in civil discourse and individual's outlook on a given expert domain. The field of artificial intelligence (AI) and machine learning (ML) are not immune to this. Public opinions of applications based on ML are informed by the interactions with such models and external opinions. However as the capacity of models increases, so to does the computational resources needed to train, run, and interact with them.This has lead to the gap between what the common user utilizes and what the cutting-edge researchers utilize to grow even further. In order to reduce this differential, there must be a reduction in computational overhead to operate these newer, higher capacity models. In this work, I demonstrate this problem in a domain application, then propose methods to reduce the computations needed in both the training and execution of neural network based applications. Then I will discuss the future areas that need to be addressed in this space, namely the problem intrinsic to all neural networks.First, I introduce my neural network based solution to the domain of off-target modeling behavior of CRISPR Cas-13 based treatments. My solution, which I call Guide-Guard, out performs previous works while being considerably smaller. By changing how the data was handled and how the model was arrange, I was able to decrease the size of the model by over 100x. This work shows the over bloating of models in domain applications.Second, I propose my solution to reduce the complexity of training titled Fine-Pruning. In this work, I demonstrate that for over bloated and generalized models can be trained over a subset of the data's features. It does this by pruning away channels based on how much they activate in response to the data. This is based on how the human brain learns. By pruning unused connections, the brain reduces noisy connections. Since this method only requires forward passes, it dramatically reduces the number of computations needed to achieve a given accuracy.Third, I address the problem of the amount of processing power needed to execute a model in my work LegoNet. In this work, I utilize clusters of blocked weight matrices to construct a code book. This allows us to record each block (or Lego) only once, and then only need the index of the Lego used. This work reduces the memory footprint of a model by 128x with minimal loss to accuracy. With LegoNet and Fine-Pruning, large models can be both trained and ran on smaller computers, making them more accessible.Lastly, I will discuss where there is further work that is needed to address the issue of trust in neural networks. Outside of being able to use them with less resources, there are further layers to what makes human designers trust the models they are using. In my concluding section, I discuss what bring a human operator into the loop can do to build trust.

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

Mode of access: World Wide Web

ISBN: 9798380845793Subjects--Topical Terms:

569006
Computer engineering.
Subjects--Index Terms:

Machine learningIndex Terms--Genre/Form:

554714
Electronic books.

Accessible Artificial Intelligence.
LDR:04613ntm a22003857 4500 001 1151819
005 20241118135831.5
006 m o d
007 cr mn ---uuuuu
008 250605s2023 xx obm 000 0 eng d
020 $a 9798380845793
035 $a (MiAaPQ)AAI30492234
035 $a AAI30492234
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Bingham, Joseph. $3 1478651
245 1 0 $a Accessible Artificial Intelligence.
264 0 $c 2023
300 $a 1 online resource (151 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: 85-05, Section: B.
500 $a Advisor: Zonouz, Saman.
502 $a Thesis (Ph.D.)--Rutgers The State University of New Jersey, School of Graduate Studies, 2023.
504 $a Includes bibliographical references
520 $a As science reaches forward with novel and greater developments, the gap between what the average individual knows that the knowledge required to understand the advanced technology that they interact with. This gap is felt in many fields, and has recently lead to a sociological phenomenon known as truth decay. Jennifer Kavanagh of RAND defines truth decay as the diminishing role that data and expert analysis effect in civil discourse and individual's outlook on a given expert domain. The field of artificial intelligence (AI) and machine learning (ML) are not immune to this. Public opinions of applications based on ML are informed by the interactions with such models and external opinions. However as the capacity of models increases, so to does the computational resources needed to train, run, and interact with them.This has lead to the gap between what the common user utilizes and what the cutting-edge researchers utilize to grow even further. In order to reduce this differential, there must be a reduction in computational overhead to operate these newer, higher capacity models. In this work, I demonstrate this problem in a domain application, then propose methods to reduce the computations needed in both the training and execution of neural network based applications. Then I will discuss the future areas that need to be addressed in this space, namely the problem intrinsic to all neural networks.First, I introduce my neural network based solution to the domain of off-target modeling behavior of CRISPR Cas-13 based treatments. My solution, which I call Guide-Guard, out performs previous works while being considerably smaller. By changing how the data was handled and how the model was arrange, I was able to decrease the size of the model by over 100x. This work shows the over bloating of models in domain applications.Second, I propose my solution to reduce the complexity of training titled Fine-Pruning. In this work, I demonstrate that for over bloated and generalized models can be trained over a subset of the data's features. It does this by pruning away channels based on how much they activate in response to the data. This is based on how the human brain learns. By pruning unused connections, the brain reduces noisy connections. Since this method only requires forward passes, it dramatically reduces the number of computations needed to achieve a given accuracy.Third, I address the problem of the amount of processing power needed to execute a model in my work LegoNet. In this work, I utilize clusters of blocked weight matrices to construct a code book. This allows us to record each block (or Lego) only once, and then only need the index of the Lego used. This work reduces the memory footprint of a model by 128x with minimal loss to accuracy. With LegoNet and Fine-Pruning, large models can be both trained and ran on smaller computers, making them more accessible.Lastly, I will discuss where there is further work that is needed to address the issue of trust in neural networks. Outside of being able to use them with less resources, there are further layers to what makes human designers trust the models they are using. In my concluding section, I discuss what bring a human operator into the loop can do to build trust.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2024
538 $a Mode of access: World Wide Web
650 4 $a Computer engineering. $3 569006
650 4 $a Computer science. $3 573171
650 4 $a Electrical engineering. $3 596380
653 $a Machine learning
653 $a Artificial intelligence
653 $a Fine-Pruning
653 $a LegoNet
655 7 $a Electronic books. $2 local $3 554714
690 $a 0464
690 $a 0984
690 $a 0544
710 2 $a ProQuest Information and Learning Co. $3 1178819
710 2 $a Rutgers The State University of New Jersey, School of Graduate Studies. $b Electrical and Computer Engineering. $3 1241232
773 0 $t Dissertations Abstracts International $g 85-05B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30492234 $z click for full text (PQDT)