東華大學圖書館 |

Language: English

Help

回圖書館首頁

手機版館藏查詢

Back

Switch To: Labeled | MARC Mode | ISBD

Dynamic Mathematics for Automated Ma...

Kuo, Nicholas I-Hsien,

Linked to FindBook

Google Book

Amazon

博客來

Dynamic Mathematics for Automated Machine Learning Techniques /

Record Type:	Electronic resources : Monograph/item
Title/Author:	Dynamic Mathematics for Automated Machine Learning Techniques // Nicholas I-Hsien Kuo.
Author:	Kuo, Nicholas I-Hsien,
Description:	1 electronic resource (248 pages)
Notes:	Source: Dissertations Abstracts International, Volume: 83-08, Section: B.
Contained By:	Dissertations Abstracts International83-08B.
Subject:	Dynamical systems. -
Online resource:	https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=28972441
ISBN:	9798780649960

Dynamic Mathematics for Automated Machine Learning Techniques /
Kuo, Nicholas I-Hsien,

Dynamic Mathematics for Automated Machine Learning Techniques /Nicholas I-Hsien Kuo. - 1 electronic resource (248 pages)

Source: Dissertations Abstracts International, Volume: 83-08, Section: B.

Data is everywhere in the modern world. From a heartwarming grin to a bulk international commercial order, our lives can be easily reflected and captured by our activities on social media platforms, street security cameras, and internet footprints. As data becomes more accessible, our society inevitably turns to data-mining out of curiosity for self-exploration and value extraction for financial competition. However and as our data becomes both quantitatively and qualitatively complicated, methods that attempt to understand data are also changing. As of the time that this thesis is written, we are currently living in a world where Neural Networks are slowly becoming the new de-facto standard for understanding industry big data.Machine learning consists classes of modelling techniques that are capable of finding patterns directly from data itself. Of its many classes, neural networks have been gaining a large amount of following as they can learn in a near-automatic fashion with minimal amount of prior knowledge on the data. This property makes neural networks highly versatile and thus they are heavily favoured in becoming the new industry standard models. Despite their popularity, several neural network methods remain opaque and the research community still struggles to fully explain their behaviours. The lack of interpretability not only halts innovation in designing better models, but it also complicates the difficulty to automate machine learning (AutoML) for neural networks to work in a dynamic environment.To this end, this thesis is devoted towards increasing the understanding of neural network properties. More specifically, we will analyse neural networks through the mathematical lens of dynamical systems in order to clarify how they transform their input data. Then, the mathematical insights will assist us in addressing three interconnected topics - recurrent neural networks, meta-learning, and continual learning - to increase the flexibility and utility of neural network models.Our first topic of discussion is recurrent neural networks (RNNs) - neural network models that are capable of handling sequential data. Advance RNNs possess an internal state, also known as the cell, to represent events that have occurred in the past. RNNs are intricate models, and the cell is recursively updated with several internal variables as the RNN receives more data. However, the variables that are employed to update the cell do not have immediately clear meanings. Thus in order to increase RNN interpretability, we compare RNNs to well-studied dynamical systems. By doing so, we demonstrate that we can establish connections between literature of machine learning and theoretical physics. This contribution also allows us to highlight the core inferential mechanism in RNN processing. It in turn enables us to propose a systematic simplification to create small-size RNNs that removes up to 85.4% secondary parameters while attains competitive performances.Our second topic concerns meta-learning, the AutoML technique of rapid knowledge acquisition. We focus on the meta-learning technique of learning to learn (L2L) which employs an RNN to update another base learner network. The RNN of L2L explores optimisation strategies superior than handcrafted alternatives to configure base learners in less time and with fewer data. However, current L2L algorithms are limited in practicality. While they are efficient, a meta-trained RNN can only update one unique base learner for one specific dataset.

English

ISBN: 9798780649960Subjects--Topical Terms:

3219613
Dynamical systems.

Dynamic Mathematics for Automated Machine Learning Techniques /
LDR:04809nmm a22003853i 4500 001 2437619
005 20251124054947.5
006 m o d
007 cr|nu||||||||
008 261011s2021 miu||||||m |||||||eng d
020 $a 9798780649960
035 $a (MiAaPQD)AAI28972441
035 $a (MiAaPQD)AustNatlU1885246708
035 $a AAI28972441
040 $a MiAaPQD $b eng $c MiAaPQD $e rda
100 1 $a Kuo, Nicholas I-Hsien, $e author. $3 3828670
245 1 0 $a Dynamic Mathematics for Automated Machine Learning Techniques / $c Nicholas I-Hsien Kuo.
264 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2021
300 $a 1 electronic resource (248 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: 83-08, Section: B.
500 $a Advisors: Harandi, Mehrtash.
502 $b Ph.D. $c The Australian National University (Australia) $d 2021.
520 $a Data is everywhere in the modern world. From a heartwarming grin to a bulk international commercial order, our lives can be easily reflected and captured by our activities on social media platforms, street security cameras, and internet footprints. As data becomes more accessible, our society inevitably turns to data-mining out of curiosity for self-exploration and value extraction for financial competition. However and as our data becomes both quantitatively and qualitatively complicated, methods that attempt to understand data are also changing. As of the time that this thesis is written, we are currently living in a world where Neural Networks are slowly becoming the new de-facto standard for understanding industry big data.Machine learning consists classes of modelling techniques that are capable of finding patterns directly from data itself. Of its many classes, neural networks have been gaining a large amount of following as they can learn in a near-automatic fashion with minimal amount of prior knowledge on the data. This property makes neural networks highly versatile and thus they are heavily favoured in becoming the new industry standard models. Despite their popularity, several neural network methods remain opaque and the research community still struggles to fully explain their behaviours. The lack of interpretability not only halts innovation in designing better models, but it also complicates the difficulty to automate machine learning (AutoML) for neural networks to work in a dynamic environment.To this end, this thesis is devoted towards increasing the understanding of neural network properties. More specifically, we will analyse neural networks through the mathematical lens of dynamical systems in order to clarify how they transform their input data. Then, the mathematical insights will assist us in addressing three interconnected topics - recurrent neural networks, meta-learning, and continual learning - to increase the flexibility and utility of neural network models.Our first topic of discussion is recurrent neural networks (RNNs) - neural network models that are capable of handling sequential data. Advance RNNs possess an internal state, also known as the cell, to represent events that have occurred in the past. RNNs are intricate models, and the cell is recursively updated with several internal variables as the RNN receives more data. However, the variables that are employed to update the cell do not have immediately clear meanings. Thus in order to increase RNN interpretability, we compare RNNs to well-studied dynamical systems. By doing so, we demonstrate that we can establish connections between literature of machine learning and theoretical physics. This contribution also allows us to highlight the core inferential mechanism in RNN processing. It in turn enables us to propose a systematic simplification to create small-size RNNs that removes up to 85.4% secondary parameters while attains competitive performances.Our second topic concerns meta-learning, the AutoML technique of rapid knowledge acquisition. We focus on the meta-learning technique of learning to learn (L2L) which employs an RNN to update another base learner network. The RNN of L2L explores optimisation strategies superior than handcrafted alternatives to configure base learners in less time and with fewer data. However, current L2L algorithms are limited in practicality. While they are efficient, a meta-trained RNN can only update one unique base learner for one specific dataset.
546 $a English
590 $a School code: 0433
650 4 $a Dynamical systems. $3 3219613
650 4 $a Neural networks. $3 677449
650 4 $a Pattern recognition. $3 3560648
650 4 $a Applied mathematics. $3 2122814
650 4 $a Computer science. $3 523869
650 4 $a Eigenvalues. $3 631789
690 $a 0800
690 $a 0984
690 $a 0364
710 2 $a The Australian National University (Australia). $e degree granting institution. $3 3765800
720 1 $a Harandi, Mehrtash $e degree supervisor.
773 0 $t Dissertations Abstracts International $g 83-08B.
790 $a 0433
791 $a Ph.D.
792 $a 2021
856 4 0 $u https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=28972441