Senior Theses - Electrical and Computer Engineering
http://hdl.handle.net/2142/30795
The best of ECE undergraduate researchThu, 21 Jun 2018 23:47:38 GMT2018-06-21T23:47:38ZA spectral method for stable bispectrum inversion with application to multireference alignment
http://hdl.handle.net/2142/100064
A spectral method for stable bispectrum inversion with application to multireference alignment
Chen, Hua
This thesis aims to develop an alignment-free method to estimate the underlying
signal from a large number of noisy randomly shifted observations,
called multi-reference alignment problem. Specifically, we make use of invariant
features including mean, power spectrum, and the bispectrum of the
signal from the observations. We propose a new algorithm using spectral
decomposition of the bispectrum phase matrix for this specific problem. For
clean signals, we show that the eigenvectors of the bispectrum phase matrix
correspond to the true phases of the signal and its shifted copies. In noisy
cases, we will select one eigenvector with largest spectral gap to estimate
the original signal. Such spectral method is robust to noise and empirically
comparable to iterative phase synchronization and optimization on phase
manifold for noise variance sigma squared less than or equal to 0.32. It can be also be used as a stable and
efficient initialization technique for local non-convex optimization for bispectrum
inversion. Using 12-fold symmetric property of bispectrum, we are able
to increase our computational efficiency by roughly ten times.
Multireference Alignment; Bispectrum
Tue, 01 May 2018 00:00:00 GMThttp://hdl.handle.net/2142/1000642018-05-01T00:00:00ZChen, HuaAtmosperic microwave plasma treatment on oil-contaminated aluminum surface
http://hdl.handle.net/2142/100063
Atmosperic microwave plasma treatment on oil-contaminated aluminum surface
Zhu, Weikun
An atmospheric microwave plasma was used to clean and activate oil-contaminated Al surfaces. Treatment was done on a large-area basis by moving the desired sample over an atmospheric pressure plasma torch. The cleaning effectiveness is supported by contact angle measurements, ATR-FTIR, SEM, XPS, and a water-break free test. In addition, the effect of processing parameters on the result of surface cleaning and activation was studied. An enhancement of surface hydrophilicity was observed as a result of increasing input microwave power and plasma exposure time. The effect of surface temperature on plasma treatment was investigated through comparing water contact angles of two methods of plasma treatment with identical exposure time, power, and gas composition. Moreover, spatially resolved ATR-FTIR spectrums from partially cleaned surfaces reveal the extent of oxidation across the plasma treated area.
Atmospheric plasma; microwave plasma; aluminum alloy surface treatment
Tue, 01 May 2018 00:00:00 GMThttp://hdl.handle.net/2142/1000632018-05-01T00:00:00ZZhu, Weikun3D human pose estimation using part affinity field
http://hdl.handle.net/2142/100062
3D human pose estimation using part affinity field
Zhao, Zixu
Nowadays, following the success of deep learning in the Computer Vision field, many research studies are underway to produce state-of-the-art technologies that can predict 3D human poses given raw image pixels. These end-to-end systems create possibilities for future studies such as human pose or gait recognition, and their practical values in industry are beyond imagination.
This thesis proposes an end-to-end system that predicts human joint locations in 3D space using only the raw image pixels as inputs. While the most used state-of-the-art method proposes that lifting joint locations from camera space to 3D space can be done in a simple and effective way only using 2D joint locations as inputs, our proposed system is even more effective and accurate with the help of part affinity fields.
Computer Vision; Pose Estimation; Neural Network
Tue, 01 May 2018 00:00:00 GMThttp://hdl.handle.net/2142/1000622018-05-01T00:00:00ZZhao, ZixuNoise analysis of a 6 muW 1.5 MHz RC relaxation oscillator
http://hdl.handle.net/2142/100061
Noise analysis of a 6 muW 1.5 MHz RC relaxation oscillator
Zhao, Peihua
With the increasing demand for small volume and low power devices, such as
the Internet of Things, RC oscillators have become more attractive. For an
RC oscillator used as a timer, the noise and jitter in an RC oscillator are vital
issues for the oscillator design. This thesis presents methods to reduce the flicker noise at low offset
frequencies of a 6 W 1.5 MHz RC oscillator with
a low dropout regulator and offset cancellation technique. This oscillator
achieves an accuracy of 30 ppmC and 1500 ppm/V. Noise in RF circuits and
the relaxation RC oscillators are analyzed first. Then two methods to reduce
oscillator noise are presented. By deleting a current mirror transistor and
adding two complementary switch pairs, the modified circuit achieves 5 time
improvement in noise performance.
RC relaxation oscillator; voltage sensitivity; regulation; flicker noise; Schmitt trigger
Tue, 01 May 2018 00:00:00 GMThttp://hdl.handle.net/2142/1000612018-05-01T00:00:00ZZhao, PeihuaAn exploration into the effect of thresholding feature maps of convolutional neural network in frequency domain
http://hdl.handle.net/2142/100060
An exploration into the effect of thresholding feature maps of convolutional neural network in frequency domain
Yu, Mang
While convolutional neural networks (CNNs) are very successful in many areas, the state-of-the-art
multi-layer CNNs usually require a large amount of computation, which limits their application in
scenarios where the computation capability is limited. Since the convolution operation can be done
efficiently in the frequency domain, researchers have successfully reduced the amount of computation
by applying the fast Fourier transform (FFT) and its inverse to the CNN. Furthermore, the sparse Fourier
transform (SFT) algorithm can further reduce the amount of computation by only extracting the salient
points in the frequency domain. However, due to this feature, it requires the inputs to be sparse or
approximately sparse in the frequency domain.
To explore the possibility of applying the SFT to CNN, we simulate the effect of SFT by removing the
frequencies with lower power. We refer to this operation as Thresholding. In the experiments, we first
inspect the effect of removing low-power frequencies for a sample feature map extracted from
intermediate outputs. The result shows that most features are still identifiable to human eyes when 90%
of the frequencies are removed; thus, it is possible that CNN can still recognize the features. We then
apply the thresholding to each individual layer of VGG-16 and test the accuracy over the ILSVRC2012
dataset. The result shows that thresholding each layer only slightly reduced the accuracy and the
reductions are smaller for the top layers (layer close to the output). However, thresholding uniformly on
every layer of the network significantly reduced the accuracy. Therefore, we conclude that we should
apply SFT to different layers with different configurations to achieve the optimal balance between
performance and accuracy. In addition, layer-by-layer fine-tuning and image processing techniques
might also help reducing the accuracy loss.
CNN; Sparse Fourier Transform
Tue, 01 May 2018 00:00:00 GMThttp://hdl.handle.net/2142/1000602018-05-01T00:00:00ZYu, MangAnalysis of the two-part predictive coder
http://hdl.handle.net/2142/100059
Analysis of the two-part predictive coder
Yu, Lian
The two-part predictive coding framework aims to compress signals while preserving feature quality for
analysis purposes. The change in feature vectors after the compression is treated as a prediction error and
is quantized using a classification centric quantizer. The classification centric quantizer is a vector quantizer
that minimizes classification error in the task of image classification. In this thesis, the method is applied
to the STL-10 dataset and a subset of the ILSVRC2012 dataset. The classification systems include a deep
hybrid neural network that consists of the scattering transform and the residual network, and an end-to-end
learned deep residual network.
Image Compression; Image Classification; Predictive Coding; Neural Networks
Tue, 01 May 2018 00:00:00 GMThttp://hdl.handle.net/2142/1000592018-05-01T00:00:00ZYu, LianModel predictive control
http://hdl.handle.net/2142/100058
Model predictive control
Yue, Haoliang
Model predictive control (MPC) refers to a family control method which applies to discrete and continuous-time process models. The future states are predicted at each time instance using the known state and system model, up to a specific time instance. The future control inputs are calculated by optimizing a determined criterion to keep the process as close as possible to the reference trajectory. An explicit solution can be obtained if the criterion is quadratic, the model is linear, and there are no constraints. Comparing to the proportional–integral–derivative (PID) feedback control, feedforward control, and inverse dynamic control that is learned in other courses, MPC has its advantages and is widely used in industry.
This thesis is done under the supervision of professor Belabbas, and the main progress is to read academic papers, books, and rederived the mathematical equations of MPC using knowledge of control and linear algebra. Derivations were implemented into MATLAB functions and the control algorithm is simulated on several different models, as well as the reaction wheel pendulum in the control lab. The state-space model is obtained in continuous time domain by LaGrange formalism, and then the equations of motion were transformed into the discrete time domain. A goal is to compare the qualitative performance of PID and MPC controller on this particular system setup.
Control System; MPC; Model Predictive Control
Tue, 01 May 2018 00:00:00 GMThttp://hdl.handle.net/2142/1000582018-05-01T00:00:00ZYue, HaoliangCausal inference for early detectection of hardware failure
http://hdl.handle.net/2142/100049
Causal inference for early detectection of hardware failure
Yang, Alan
Many modern hardware systems are equipped with sensors which record
time-series diagnostic data. These sensors enable data-driven failure prediction
that can reduce the need for component redundancy and lengthen
lifetime specifications, by allowing for identification and proactive replacement
of a soon-to-fail component. In this work, we develop a causal inference
framework for predicting data center hard disk drive failures using
multivariate time series recordings of temperature, read error rate, and other
attributes. Information-theoretic measures are developed to quantify relationships
between sensor variables, select prognostic features, and train a
predictor. Finally, a recurrent neural network demonstrating high predictive
accuracy and a low false alarm rate is developed, using field data collected
from an operating data center.
hardware failure detection; predicting hard disk drive failures; information theoretic measures
Tue, 01 May 2018 00:00:00 GMThttp://hdl.handle.net/2142/1000492018-05-01T00:00:00ZYang, AlanArt painting author and style labeling using convolutional neural network
http://hdl.handle.net/2142/100048
Art painting author and style labeling using convolutional neural network
Xu, Ke
This thesis proposes a convolutional neural network-based approach for labeling art paintings by their author and style. Our Artist1000 dataset consists of 1000 raw painting images for each of 5 prolific artists. Our Style1000 dataset consists of 1000 raw painting images for each of 5 different painting styles. Two independent ResNet18 classifiers are trained and validated, one for each dataset. They are combined to predict the author and style of unseen art paintings. Our algorithm first classifies patches extracted from raw painting images and then uses majority vote of patches from the same image to predict image label. We achieve 84.3% accuracy on Artist1000 dataset and 66.9% accuracy on Style1000 dataset. Compared with the traditional methods of art painting labeling, which heavily depended on extracting complex artificial features from raw images, our algorithm shows promising results of empowering a neural network to extract local features and make predictions.
art painting; labeling; convolutional neural network; ResNet; Image sampling; Image databases
Tue, 01 May 2018 00:00:00 GMThttp://hdl.handle.net/2142/1000482018-05-01T00:00:00ZXu, KeUsing Wasserstein GAN to generate high quality adversarial examples
http://hdl.handle.net/2142/100047
Using Wasserstein GAN to generate high quality adversarial examples
Xiong, Zhihan
Although Deep Neural Networks (DNNs) have state-of-the-art performance
in various machine learning tasks, in recent years, they are found to be
vulnerable to so-called adversarial examples Specifically, take x is an element of D on
which a neural network has very high classification accuracy. It is possible to
find some small perturbation Δx so that even though the difference between
x and x + Δx = x′ is almost imperceptible to humans, the given neural
network is very likely to incorrectly classify x + Δx.
Several gradient and optimization based methods have been proposed to
create such adversarial examples x′, but many of them cannot achieve high
speed and high quality x′ simultaneously. In this thesis, we propose a new
algorithm to generate adversarial examples based on Generative Adversarial
Networks (GANs), specifically, a modification to the training algorithm of
the Improved Wasserstein GAN. The trained generator is able to create x′
very similar to the original x while keeping the classification accuracy of the
target model as low as the state-of-the-art attack. Furthermore, although
training a GAN might be slow, after it is trained, it can generate adversarial
examples much faster than previous optimization-based methods. Our goal
is for this work to be used for further research on robust neural networks.
adversarial machine learning; white-box targeted attack; Wasserstein GAN; neural networks
Tue, 01 May 2018 00:00:00 GMThttp://hdl.handle.net/2142/1000472018-05-01T00:00:00ZXiong, Zhihan