In-materio reservoir computing with single-walled carbon nanotube/porphyrin-polyoxometalate random network —A novel approach for next-generation machine intelligence

Deep,  Banerjee

WEKO3

lat lon distance

[[sub_check.contents]]

[[sub_radio.contents]]

Field does not validate

[[sub_attr.contents]]　

インデックスツリー

アイテム

単層カーボンナノチューブ/ポルフィリン-ポリ酸ランダムネットワークを用いたマテリアルリザバー演算素子 —次世代機械知能への新規アプローチ

https://doi.org/10.18997/00008819

名前 / ファイル	ライセンス	アクション
sei_k_425.pdf (8.0 MB)

アイテムタイプ

学位論文 = Thesis or Dissertation(1)

公開日

2022-04-22

資源タイプ

資源タイプ識別子

http://purl.org/coar/resource_type/c_db06

資源タイプ

doctoral thesis

タイトル

In-materio reservoir computing with single-walled carbon nanotube/porphyrin-polyoxometalate random network —A novel approach for next-generation machine intelligence

言語

タイトル

単層カーボンナノチューブ/ポルフィリン-ポリ酸ランダムネットワークを用いたマテリアルリザバー演算素子 —次世代機械知能への新規アプローチ

言語

eng

著者

Deep, Banerjee

抄録

内容記述タイプ

Abstract

内容記述

In a layman’s term, computation is defined as the execution of a given instruction through a programmable algorithm. History has it that starting from the simplest calculator to the sophisticated von Neumann machine, the above definition has been followed without a flaw. Logical operations for which a human takes a minute long to solve, is a matter of fraction of seconds for these gadgets. But contrastingly, when it comes to critical and analytical thinking that requires learning through observation like the human brain, these powerful machines falter and lag behind. Thus, inspired from the brain’s neural circuit, software models of neural networks （NN） integrated with high-speed supercomputers were developed as an alternative tool to implement machine intelligent tasks of function optimization, pattern, and voice recognition. But as device downscaling and transistor performance approaches the constant regime of Moore’s law due to high CMOS fabrication cost and large tunneling energy loss, training these algorithms over multiple hidden layers is turning out to be a grave concern for future applications. As a result, the interplay between faster performance and low computational power requirement for complex tasks deems highly disproportional. Therefore, alternative in terms of both NN models and conventional Neumann architecture needs to be addressed in today’s age for next-generation machine intelligence systems. Fortunately, through extensive research and studies, unconventional computing using a reservoir based neural network platform, called in-materio reservoir computing （RC） has come to the rescue. In-maerio RC uses physical, biological, chemical, cellular automata and other inanimate dynamical systems as a source of non-linear high dimensional spatio-temporal information processing unit to construct a specific target task. RC not only has a three-layer simplified neural architectural layer, but also imposes a cheap, fast, and simplified optimization of only the readout weights with machine intelligent regression algorithm to construct the supervised objective target via a weighted linear combination of the readouts. Thus, utilizing this idea, herein in this work we report such an in-materio RC with a dynamical random network of single walled carbon nanotube/porphyrin-polyoxometalate （SWNT/Por-POM） device. We begin with Chapter 1, which deals with the introduction covering the literature of ANN evolution and the shortcomings of von Neumann architecture and training models of these ANN, which leads us to adopt the in-materio RC architecture. We design the problem statement focused on extending the theoretical RC model of previously suggested SWNT/POM network to an experimental one and present the objective of fabricating a random network based on nanomaterials as they closely resemble the network structure of the brain. Finally, we conclude by stating the scope of this research work aiming towards validating the non-linear high dimensional reservoir property SWNT/Por-POM holds for it to explicitly demonstrate the RC benchmark tasks of optimization and classification. Chapter 2 describes the methodology including the chemical repository required for the facile synthesis of the material. The synthesis part is divided broadly into SWNT purification and then its dispersion with Por-POM to form the desired complex. It is then followed up with the microelectrode array fabrication and the consequent wet-transfer thin film deposition to give the ultimate reservoir architecture of input-output control read pads with SWNT/Por-POM reservoir. Finally we give a briefing of AFM, UV-Vis spectroscopy, FE-SEM characterization techniques of SWNT/Por-POM complex along with the electrical set-up interfaced with software algorithm to demonstrate the RC approach of in-materio machine intelligence. In Chapter 3, we study the current dynamics as a function of voltage and time and validate the non-linear information processing ability intrinsic to the device. The study reveals that the negative differential resistance （NDR） arising from redox nature of Por-POM results in oscillating random noise outputs giving rise to 1/f brain-like spatio-temporal information. We compute the memory capacity （MC） and prove that the device exhibits echo state property of fading memory, but remembers very little of the past information. The low MC and high non-linearity allowed us to choose mostly non-linear tasks of waveform generation, Boolean logic optimization and one-hot vector binary object classification as the RC benchmark. The Chapter 4 relates to the waveform generation task. Utilizing the high dimensional voltage readouts of varying amplitude, phase and higher harmonic frequencies, relative to input sine wave, a regression optimization was performed towards constructing cosine, triangular, square and sawtooth waves resulting in a high accuracy of around 95%. The task complexity of function optimization was further enhanced in Chapter 5 where two inputs were used to construct Boolean logic functions of OR, AND, XOR, NOR, NAND and XNOR. Similar to the waveform, accuracy over 95% could be achieved due to the presence of NDR nonlinearity. Furthermore, the device was also tested for classification problem in Chapter 6. Here we showed an off-line binary classification of four object toys; hedgehog, dog, block and bus, using the grasped tactile information of these objects as inputs obtained from the Toyota Human Support Robot. A one-ridge regression analysis to fit the hot vector supervised target was used to optimize the output weights for predicting the correct outcome. All the objects were successfully classified owing to the 1/f information processing factor. Lastly, we conclude the section in Chapter 7 with the future scope of extending the idea to fabricate a 3-D model of the same material as it opens up opportunity for higher memory capacity fruitful for future benchmark tasks of time-series prediction. Overall, our research marks a step stone in utilizing SWNT/Por-POM as the in-materio RC for the very first time thereby making it a desirable candidate for next-generation machine intelligence.

内容記述タイプ

TableOfContents

内容記述

1 Introduction and Literature review||2 Methodology||3 Reservoir dynamics emerging from an incidental structure of single-walled carbon nanotube/porphyrin-polyoxometalate complex||4 Fourier transform waveforms via in-materio reservoir computing from single-walled carbon nanotube/porphyrin-polyoxometalate complex||5 Room temperature demonstration of in-materio reservoir computing for optimizing Boolean function with single-walled carbon nanotube/porphyrin-polyoxometalate composite||6 Binary object classification with tactile sensory input information of via single-walled carbon nanotube/porphyrin-polyoxometalate network as in-materio reservoir computing||7 Future scope and Conclusion

備考

内容記述タイプ

Other

内容記述

九州工業大学博士学位論文学位記番号：生工博甲第425号学位授与年月日：令和3年12月27日

キーワード

主題Scheme

Other

主題

Single walled carbon nanotube

キーワード

主題Scheme

Other

主題

Porphyrin polyoxometalate

キーワード

主題Scheme

Other

主題

Reservoir computing

キーワード

主題Scheme

Other

主題

Physical device

キーワード

主題Scheme

Other

主題

Machine intelligence

アドバイザー

田中, 啓文

学位授与番号

甲第425号

学位名

博士（工学）

学位授与年月日

2021-12-27

学位授与機関

学位授与機関識別子Scheme

kakenhi

学位授与機関識別子

17104

学位授与機関名

九州工業大学

学位授与年度

内容記述タイプ

Other

内容記述

令和3年度

出版タイプ

VoR

出版タイプResource

http://purl.org/coar/version/c_970fb48d4fbd8a85

アクセス権

open access

アクセス権URI

http://purl.org/coar/access_right/c_abf2

ID登録

10.18997/00008819

ID登録タイプ

JaLC

戻る

views

See details

	Views

Versions

Ver.1

2023-05-15 12:56:31.226133

Show All versions

Cite as

Other

エクスポート

OAI-PMH

JPCOAR 2.0
JPCOAR 1.0
DublinCore
DDI

Other Formats

インデックスリンク

インデックスツリー

アイテム

単層カーボンナノチューブ/ポルフィリン-ポリ酸ランダムネットワークを用いたマテリアルリザバー演算素子 —次世代機械知能への新規アプローチ

× Deep, Banerjee

Versions

Share

Cite as

Other

エクスポート

コミュニティ

メニューを最小化

インデックスリンク

インデックスツリー

アイテム

単層カーボンナノチューブ/ポルフィリン-ポリ酸ランダムネットワークを用いたマテリアルリザバー演算素子 —次世代機械知能への新規アプローチ

× Deep, Banerjee

Versions

Share

Cite as

Other

エクスポート

コミュニティ