Hasil Pencarian  ::  Simpan CSV :: Kembali

"The goal of this book is a complete framework for classifying and transcribing sequential data with recurrent neural networks only. Three main innovations are introduced in order to realise this goal. Firstly, the connectionist temporal classification output layer allows the framework to be trained with unsegmented target sequences, such as phoneme-level speech transcriptions, this is in contrast to previous connectionist approaches, which were dependent on error-prone prior segmentation. Secondly, multidimensional recurrent neural networks extend the framework in a natural way to data with more than one spatio-temporal dimension, such as images and videos. Thirdly, the use of hierarchical subsampling makes it feasible to apply the framework to very large or high resolution sequences, such as raw audio or video."

"The second edition of the book augments the first edition with more tricks, which have resulted from 14 years of theory and experimentation by some of the world's most prominent neural network researchers. These tricks can make a substantial difference (in terms of speed, ease of implementation, and accuracy) when it comes to putting algorithms to work on real problems."

"Kemajuan teknologi mengiringi kemajuan Pesawat Tanpa Awak yang membuat peneliti terus mengembangkannya. Quadcopter merupakan Pesawat Tanpa Awak yang saat ini telah banyak digunakan untuk berbagai tujuan. Bentuknya yang ringkas serta beratnya yang ringan dengan empat buah baling-baling motor membuat quadcopter memiliki keunggulan dalam kemampuan dalam melakukan maneuver di udara. Tujuan dari penelitian skripsi ini adalah diajukannya sebuah ide menciptakam pengendali Jaringan Saraf Kendali Inverse Langsung NN ndash;DIC ndash; Neural Network Direct Inverse Control dengan algoritma Elman Recurrent untuk quadcopter, dan membandingkannya dengan pengendali berbasis algoritma Back Propagation Neural Network biasa. Dalam skripsi ini dikemukakan hasil simulasi dari identifikasi quadcopter dengan memodelkan secara black box, serta hasil dari dua jenis pengendali Inverse untuk quadcopter yaitu Elman Recurrent Neural Network Direct Inverse Control dan Back Propagation Neural Network Direct Inverse Control.

---

Technological advances accompany the progress of Unmanned Aircraft that keeps researchers on the rise. Quadcopter is an Unmanned Aircraft that is now widely used for various purposes. Its compact shape and light weight with four motor propellers make the quadcopter has an advantage in the ability to maneuver in the air. The purpose of this thesis research is to propose an idea to create a controller of the Direct Inverse Control Neural Network NN ndash DIC with Elman Recurrent algorithm for quadcopter, and compare it with an ordinary Back Propagation Neural Netwok algorithm. In this thesis, the shown simulation results are those of quadcopter plant based on black box modeling identification, and the result of two types of Inverse controllers for quadcopter, Elman Recurrent Neural Netwok Direct Inverse Control and Back Propagation Neural Network Direct Inverse Control."

"Debit air (flowrate) yang bernilai konstan merupakan besaran fisis fundamental dalam sistem transportasi fluida dari satu tempat ke tempat lain. Untuk mencapai hal tersebut, dibutuhkan suatu sistem kendali yang mampu menghasilkan debit air bernilai konstan. Pada penelitian ini, transportasi fluida dibuat dalam sebuah rangkaian plant miniatur dengan menerapkan sistem kendali didalamnya. Pada plant tersebut terdapat actuator control valve, flow transmitter, dan Programmable Logic Controller (PLC). OPC Server merupakan perangkat lunak antarmuka menggunakan mode client/server berbasis COM/ DCOM yang memungkinkan MATLAB dapat berkomunikasi dengan PLC. Dalam proses komunikasi antara PLC dengan MATLAB digunakan OPC server yang berfungsi sebagai "jembatan" antara keduanya. Sistem kendali yang diterapkan berupa PID-Controller dan soft computing Neural Network (NN) dengan menggunakan MATLAB SIMULINK. Penerapan soft computing Neural Network (NN) bertujuan untuk mengoptimasi performa sistem kendali PID-Controller yang telah umum digunakan. Faktor-faktor performa yang dijadikan parameter pembanding adalah nilai rise time, settling time, maximum overshoot, dan steady-state error. Berdasarkan hasil percobaan, Neural Network Controller memiliki nilai permformansi yang lebih baik daripada PID-Controller. Nilai performansi Neural Network Controller yang didapatkan yakni maximum overshoot = 5.36% dan steady-state error = 0.85%.

---

Flowrate is a fundamental physical quantity in the fluid transportation system from one place to another. To achieve this, a control system is needed that is able to produce a constant flow of water. In this study, fluid transport was made in a miniature plant series by implementing a control system in it. At the plant there is a control valve actuator, flow transmitter, and Programmable Logic Controller (PLC). OPC Server is interface software using COM / DCOM-based client / server mode that allows MATLAB to communicate with the PLC. In the process of communication between PLC and MATLAB, the OPC server is used as a "bridge" between the two. The control system applied is in the form of PID-Controller and soft computing Neural Network (NN) using MATLAB SIMULINK. The application of soft computing Neural Network (NN) aims to optimize the performance of the PID-Controller control system that has been commonly used. Performance factors that are used as comparison parameters are the value of rise time, settling time, maximum overshoot, and steady-state error. Based on the results of the experiment, the Neural Network Controller has a better value of permformance than PID-Controller. The performance value of the Neural Network Controller obtained is maximum overshoot = 5.36% and steady-state error = 0.85%."

"ABSTRAKNeural networks merupakan salah satu pendekatan yang sering digunakan dalam melakukan analisis data. Dalam perkembangannya, neural networks mencapai kesuksesan dalam berbagai bidang, mulai dari pengenalan gambar, representasi bahasa,hingga bio informatika. Beberapa penelitian terakhir menunjukkan bahwa model neural networks memiliki kekurangan dalam melindungi informasi yang terdapat dalam training set agar tidak dapat dieksploitasi oleh pihak-pihak yang tidak berkepentingan. Kekurangan ini dapat dieksploitasi dengan membuat sebuah model yang dapat menentukan apakah seseorang berada dalam training set atau tidak, dan hasilnya dapat digunakan untuk melanggar privasi orang tersebut. Eksploitasi ini disebut dengan serangan membership inference. Serangan membership infrerence dapat dihindari oleh model yang memenuhi kriteria differential privacy, yaitu probabilitas keluaran dari model pada dua database yang berbeda pada satu baris pada dasarnya mirip. Pada tesis ini, dikembangkan algoritma optimisasi berbasis gradien seperti Momentum, Nesterov, RMSProp dan Adam yang memenuhi kriteria differential privacy. Algoritma yang dikembangkan digunakan untuk melatih model neural networks agar memenuhi kriteria differential privacy. Eksperimen yang dilakukan menunjukkan bahwa algoritma yang dikembangkan dapat digunakan untuk melatih model neural networks dan menghasilkan model yang lebih akurat dibandingkan algoritma stochastic gradient descent yang memenuhi kriteria differential privacy. Diperlihatkan juga pengaruh penjaminan privasi terhadap akurasi model yang dilatih menggunakan algoritma yang dikembangkan, yaitu penjaminan privasi yang lebih kuat menghasilkan akurasi model yang lebih rendah, dan sebaliknya.

---

ABSTRACT

Neural networks is one of the popular approach to analyze data. It has showed excellent ability to tackle complex problems in various domain, e.g., computer vision,language representation, and bioinformatics. At some point, neural network model may leak some information about the training data. This leakage could be exploited by adversaries to violate individuals in the training data. Membership inference attack is one kind of attacks that could be used by the adversary. This attack can be mitigated by using differentially private models. In this thesis, differentially private optimization algorithms, i.e., momentum, nesterov, rmsprop, adam, were developed. These algorithms then used to train a differentially private neural networks model. It was shown by the experiments conducted that these algorithms can be used to train a neural networks model, and yields better model accuracy compared to stochastic gradient descent algorithm. The tradeoff between privacy and utility is also studied."

"Instructs graduate and undergraduate-level students in electrical engineering, informatics, control engineering, mechanics, robotics, bioengineering on the concepts of complex-valued neural networks. This title focuses on neural networks that deal with complex numbers and the practical advantages of complex-valued neural networks. "