Digital Commons Network^™

Systematic Comparison Of Reservoir Computing Frameworks, Nihar S. Koppolu, Christof Teuscher

Student Research Symposium

In this poster, we present a systematic evaluation and comparison of five Reservoir computing (RC) software simulation frameworks, namely reservoirpy, RcTorch, pyRCN, pytorch-esn, and ReservoirComputing.jl. RC is a specific machine learning approach that leverages fixed, nonlinear systems to map signals into higher dimensions. Its unique strength lies in training only the readout layer, which reduces the training complexity. RC excels in temporal signal processing and is also well suited for various physical implementations. The increasing interest in RC has led to the proliferation of various RC simulation frameworks. Our RC simulation framework evaluation focuses on a feature comparison, documentation quality, …

Energy-Efficient Neuromorphic Architectures For Nuclear Radiation Detection Applications, Jorge I. Canales-Verdial, Jamison R. Wagner, Landon A. Schmucker, Mark Wetzel, Nathan J. Withers, Philippe Erol Proctor, Christof Teuscher, Multiple Additional Authors

Electrical and Computer Engineering Faculty Publications and Presentations

A comprehensive analysis and simulation of two memristor-based neuromorphic architectures for nuclear radiation detection is presented. Both scalable architectures retrofit a locally competitive algorithm to solve overcomplete sparse approximation problems by harnessing memristor crossbar execution of vector–matrix multiplications. The proposed systems demonstrate excellent accuracy and throughput while consuming minimal energy for radionuclide detection. To ensure that the simulation results of our proposed hardware are realistic, the memristor parameters are chosen from our own fabricated memristor devices. Based on these results, we conclude that memristor-based computing is the preeminent technology for a radiation detection platform.

Toward Efficient Rendering: A Neural Network Approach, Qiqi Hou

Dissertations and Theses

Physically-based image synthesis has attracted considerable attention due to its wide applications in visual effects, video games, design visualization, and simulation. However, obtaining visually satisfactory renderings with ray tracing algorithms often requires casting a large number of rays and thus takes a vast amount of computation. The extensive computational and memory requirements of ray tracing methods pose a challenge, especially when running these rendering algorithms on resource-constrained platforms, and impede their applications that require high resolutions and refresh rates. This thesis presents three methods to address the challenge of efficient rendering.

First, we present a hybrid rendering method to speed …

Sequential Frame-Interpolation And Dct-Based Video Compression Framework, Yeganeh Jalalpour, Wu-Chi Feng, Feng Liu

Computer Science Faculty Publications and Presentations

Video data is ubiquitous; capturing, transferring, and storing even compressed video data is challenging because it requires substantial resources. With the large amount of video traffic being transmitted on the internet, any improvement in compressing such data, even small, can drastically impact resource consumption. In this paper, we present a hybrid video compression framework that unites the advantages of both DCT-based and interpolation-based video compression methods in a single framework. We show that our work can deliver the same visual quality or, in some cases, improve visual quality while reducing the bandwidth by 10--20%.

Unpaired Style Transfer Conditional Generative Adversarial Network For Scanned Document Generation, David Jonathan Hawbaker

Dissertations and Theses

Neural networks are a powerful machine learning tool, especially when trained on a large dataset of relevant high-quality data. Generative adversarial networks, image super resolution and most other image manipulation neural networks require a dataset of images and matching target images for training. Collecting and compiling that data can be time consuming and expensive. This work explores an approach for building a dataset of paired document images with a matching scanned version of each document without physical printers or scanners. A dataset of these document image pairs could be used to train a generative adversarial network or image super resolution …

Efficient Neuromorphic Algorithms For Gamma-Ray Spectrum Denoising And Radionuclide Identification, Merlin Phillip Carson

Dissertations and Theses

Radionuclide detection and identification are important tasks for deterring a potentially catastrophic nuclear event. Due to high levels of background radiation from both terrestrial and extraterrestrial sources, some form of noise reduction pre-processing is required for a gamma-ray spectrum prior to being analyzed by an identification algorithm so as to determine the identity of anomalous sources. This research focuses on the use of neuromorphic algorithms for the purpose of developing low power, accurate radionuclide identification devices that can filter out non-anomalous background radiation and other artifacts created by gamma-ray detector measurement equipment, along with identifying clandestine, radioactive material.

A sparse …

Proximal Policy Optimization For Radiation Source Search, Philippe Erol Proctor

Dissertations and Theses

Rapid localization and search for lost nuclear sources in a given area of interest is an important task for the safety of society and the reduction of human harm. Detection, localization and identification are based upon the measured gamma radiation spectrum from a radiation detector. The nonlinear relationship of electromagnetic wave propagation paired with the probabilistic nature of gamma ray emission and background radiation from the environment leads to ambiguity in the estimation of a source's location. In the case of a single mobile detector, there are numerous challenges to overcome such as weak source activity, multiple sources, or the …

On The (Im)Practicality Of Adversarial Perturbation For Image Privacy, Arezoo Rajabi, Rakesh B. Bobba, Mike Rosulek, Charles Wright, Wu-Chi Feng

Computer Science Faculty Publications and Presentations

Image hosting platforms are a popular way to store and share images with family members and friends. However, such platforms typically have full access to images raising privacy concerns. These concerns are further exacerbated with the advent of Convolutional Neural Networks (CNNs) that can be trained on available images to automatically detect and recognize faces with high accuracy.

Recently, adversarial perturbations have been proposed as a potential defense against automated recognition and classification of images by CNNs. In this paper, we explore the practicality of adversarial perturbation based approaches as a privacy defense against automated face recognition. Specifically, we first …

View Synthesis Of Dynamic Scenes Based On Deep 3d Mask Volume, Kai-En Lin, Guowei Yang, Lei Xiao, Feng Liu, Ravi Ramamoorthi

Computer Science Faculty Publications and Presentations

Image view synthesis has seen great success in reconstructing photorealistic visuals, thanks to deep learning and various novel representations. The next key step in immersive virtual experiences is view synthesis of dynamic scenes. However, several challenges exist due to the lack of high-quality training datasets, and the additional time dimension for videos of dynamic scenes. To address this issue, we introduce a multi-view video dataset, captured with a custom 10-camera rig in 120FPS. The dataset contains 96 high-quality scenes showing various visual effects and human interactions in outdoor scenes. We develop a new algorithm, Deep 3D Mask Volume, which enables …

Vessel Trajectory Prediction Using Historical Ais Data, Jagir Laxmichand Charla

Dissertations and Theses

Maritime vessel position coordinates are important information for maritime situational planning and organization. A better estimate of future locations of the maritime vessels, from their current locations, can help maritime authorities to make planned decisions, which can be helpful to avoid traffic congestion and longer waiting times. This thesis develops a method for estimating future locations of the vessels using their current and previous locations and other data.

The motivating scenario for this work is that of determining the future locations of the vessels based on their current location and previous locations for accurate modelling of underwater acoustic noise. As …

Extending The Functional Subnetwork Approach To A Generalized Linear Integrate-And-Fire Neuron Model, Nicholas Szczecinski, Roger Quinn, Alexander J. Hunt

Mechanical and Materials Engineering Faculty Publications and Presentations

Engineering neural networks to perform specific tasks often represents a monumental challenge in determining network architecture and parameter values. In this work, we extend our previously-developed method for tuning networks of non-spiking neurons, the “Functional subnetwork approach” (FSA), to the tuning of networks composed of spiking neurons. This extension enables the direct assembly and tuning of networks of spiking neurons and synapses based on the network’s intended function, without the use of global optimization ormachine learning. To extend the FSA, we show that the dynamics of a generalized linear integrate and fire (GLIF) neuronmodel have fundamental similarities to those of …

Exploring The Potential Of Sparse Coding For Machine Learning, Sheng Yang Lundquist

Dissertations and Theses

While deep learning has proven to be successful for various tasks in the field of computer vision, there are several limitations of deep-learning models when compared to human performance. Specifically, human vision is largely robust to noise and distortions, whereas deep learning performance tends to be brittle to modifications of test images, including being susceptible to adversarial examples. Additionally, deep-learning methods typically require very large collections of training examples for good performance on a task, whereas humans can learn to perform the same task with a much smaller number of training examples.

In this dissertation, I investigate whether the use …

Flight Simulator Modeling Using Recurrent Neural Networks, Nickolas Sabatini, Andreas Natsis

Undergraduate Research & Mentoring Program

Recurrent neural networks (RNNs) are a form of machine learning used to predict future values. This project uses RNNs tor predict future values for a flight simulator. Coded in Python using the Keras library, the model demonstrates training loss and validation loss, referring to the error when training the model.

Fallen Objects: Collaborating With Artificial Intelligence In The Field Of Graphic Design, Harrison S. Gerard

University Honors Theses

In this paper, I discuss the creation, execution and reception of my digital art series Fallen Objects, in which I collaborate with a neural net to create pseudo-found objects. I explore how artists might collaborate with Artificial Intelligence obliquely, not by having the AI generate the images themselves, but instead generate input for the artists to make the images. While many artists are focused on training neural nets to replicate their own art inputs, I instead focus on working with an AI trained on external, easily-accessible data and creating images from the prompts it delivers. In this way, the AI …

Minimum Complexity Echo State Networks For Genome And Sequence Analysis, Christopher John Neighbor

Dissertations and Theses

Increasing viral illnesses threatens global human health and welfare. Due to the distribution of disease and the expense of diagnosis, it is of value to develop portable assays that can detect viral infections early. DNA molecular logic technology offers a portable detection method due to the versatility and stability of DNA and the potential of in situ computation.

Top-down engineering of these chemical logic networks can be difficult due to the difficulties of their implementation using DNA as a substrate. In this work echo state networks, a form of recurrent neural networks, were explored with the motivation that their implementation …

Experimenting With A Biologically Plausible Neural Network, Dmitri Murphy

University Honors Theses

We present research on an implementation of a biologically inspired Bayesian Confidence Propagation Neural Network (BCPNN). Based on previous work by Christopher Johansson and Anders Lansner, our implementation seeks to test and understand the various properties of this model. The floating-point implementation we built uses discrete time and bit-vectors as input/output. We found that the column based BCPNN model is able to memorize a decent number of input vectors and is able to restore noisy versions of these vectors with relatively high accuracy. We examine the model’s capacity, noise recovery ability and cross-column connection influence, among other attributes. The clearest …

An Application Of Deep Learning Models To Automate Food Waste Classification, Alejandro Zachary Espinoza

Dissertations and Theses

Food wastage is a problem that affects all demographics and regions of the world. Each year, approximately one-third of food produced for human consumption is thrown away. In an effort to track and reduce food waste in the commercial sector, some companies utilize third party devices which collect data to analyze individual contributions to the global problem. These devices track the type of food wasted (such as vegetables, fruit, boneless chicken, pasta) along with the weight. Some devices also allow the user to leave the food in a kitchen container while it is weighed, so the container weight must also …

Sensory Relevance Models, Walt Woods

Dissertations and Theses

This dissertation concerns methods for improving the reliability and quality of explanations for decisions based on Neural Networks (NNs). NNs are increasingly part of state-of-the-art solutions for a broad range of fields, including biomedical, logistics, user-recommendation engines, defense, and self-driving vehicles. While NNs form the backbone of these solutions, they are often viewed as "black box" solutions, meaning the only output offered is a final decision, with no insight into how or why that particular decision was made. For high-stakes fields, such as biomedical, where lives are at risk, it is often more important to be able to explain a …

Design Of A Canine Inspired Quadruped Robot As A Platform For Synthetic Neural Network Control, Cody Warren Scharzenberger

Dissertations and Theses

Legged locomotion is a feat ubiquitous throughout the animal kingdom, but modern robots still fall far short of similar achievements. This paper presents the design of a canine-inspired quadruped robot named DoggyDeux as a platform for synthetic neural network (SNN) research that may be one avenue for robots to attain animal-like agility and adaptability. DoggyDeux features a fully 3D printed frame, 24 braided pneumatic actuators (BPAs) that drive four 3-DOF limbs in antagonistic extensor-flexor pairs, and an electrical system that allows it to respond to commands from a SNN comprised of central pattern generators (CPGs). Compared to the previous version …

The Improvement Of Machine Learning Accuracies Through Transfer Learning, Jordan T. Le

University Honors Theses

Pretrained models could be reused in a way that allows for improvement in training accuracy. Training a model from scratch takes time. The goal is improving accuracy and minimizing the loss across individual epochs. The hypothesis is that transfer learning could potentially improve on the rate of accuracy and speed of training per epoch iteration.

An Evaluation Of Vgg16 And Yolo V3 On Hand-Drawn Images, Lee Hoang

University Honors Theses

This thesis evaluates the accuracy and performance of VGG16, a convolutional neural network (CNN), and YOLO v3, an object detector, on a dataset of 1000 hand-drawn images. Unlike with photographs, which possess high amounts of detail, sketches tend to lack much detail aside from the freehand lines that comprise them. This is further detailed in prior works about Sketch-based Image Retrieval (SBIR), a classification task to map photographs to sketches; and SketchParse, a CNN that analyzes sketch features and assigns captions. In this paper, I show the differences in classification accuracy between VGG16 and YOLO v3. The former model, pretrained …

Exploring And Expanding The One-Pixel Attack, Umairullah Khan, Walt Woods, Christof Teuscher

Student Research Symposium

In machine learning research, adversarial examples are normal inputs to a classifier that have been specifically perturbed to cause the model to misclassify the input. These perturbations rarely affect the human readability of an input, even though the model’s output is drastically different. Recent work has demonstrated that image-classifying deep neural networks (DNNs) can be reliably fooled with the modification of a single pixel in the input image, without knowledge of a DNN’s internal parameters. This “one-pixel attack” utilizes an iterative evolutionary optimizer known as differential evolution (DE) to find the most effective pixel to perturb, via the evaluation of …

The Applications Of Grid Cells In Computer Vision, Keaton Kraiger

Undergraduate Research & Mentoring Program

In this study we present a novel method for position and scale invariant object representation based on a biologically-inspired framework. Grid cells are neurons in the entorhinal cortex whose multiple firing locations form a periodic triangular array, tiling the surface of an animal’s environment. We propose a model for simple object representation that maintains position and scale invariance, in which grid maps capture the fundamental structure and features of an object. The model provides a mechanism for identifying feature locations in a Cartesian plane and vectors between object features encoded by grid cells. It is shown that key object features …

Localizing Little Landmarks With Transfer Learning, Sharad Kumar

Dissertations and Theses

Locating a small object in an image -- like a mouse on a computer desk or the door handle of a car -- is an important computer vision problem to solve because in many real life situations a small object may be the first thing that gets operated upon in the image scene. While a significant amount of artificial intelligence and machine learning research has focused on localizing prominent objects in an image, the area of small object detection has remained less explored. In my research I explore the possibility of using context information to localize small objects in an …

A Dendritic Transfer Function In A Novel Fully-Connected Layer, Mark Robert Musil

University Honors Theses

Dendritic branch operations in pyramidal neurons are well understood in-vivo but their potential as computational assets in deep neural networks has not been explored. The pre-processing which dendrites perform may be able to decrease the error of an artificial neuron because each dendrite serves as an independent filtering mechanism which may prevent false positives. In order to test this hypothesis, a fully-connected layer implementing the dendritic transfer function is defined and used to replace the final fully-connected layer used in a standard CNN (convolutional neural network). Results show that the defined algorithm is not able to predict better than chance …

Artificial Intelligence Hits The Barrier Of Meaning, Melanie Mitchell

Computer Science Faculty Publications and Presentations

Today’s AI systems sorely lack the essence of human intelligence: Understanding the situations we experience, being able to grasp their meaning. The lack of humanlike understanding in machines is underscored by recent studies demonstrating lack of robustness of state-of-the-art deep-learning systems. Deeper networks and larger datasets alone are not likely to unlock AI’s “barrier of meaning”; instead the field will need to embrace its original roots as an interdisciplinary science of intelligence.

Exploring And Expanding The One-Pixel Attack, Umairullah Khan, Walt Woods

Undergraduate Research & Mentoring Program

In machine learning research, adversarial examples are normal inputs to a classifier that have been specifically perturbed to cause the model to misclassify the input. These perturbations rarely affect the human readability of an input, even though the model’s output is drastically different. Recent work has demonstrated that image-classifying deep neural networks (DNNs) can be reliably fooled with the modification of a single pixel in the input image, without knowledge of a DNN’s internal parameters. This “one-pixel attack” utilizes an iterative evolutionary optimizer known as differential evolution (DE) to find the most effective pixel to perturb, via the evaluation of …

Emulating Balance Control Observed In Human Test Subjects With A Neural Network, Wade William Hilts

Dissertations and Theses

Human balance control is a complex feedback system that must be adaptable and robust in an infinitely varying external environment. It is probable that there are many concurrent control loops occurring in the central nervous system that achieve stability for a variety of postural perturbations. Though many engineering models of human balance control have been tested, no models of how these controllers might operate within the nervous system have yet been developed. We have focused on building a model of a proprioceptive feedback loop with simulated neurons. The proprioceptive referenced portion of human balance control has been successfully modeled by …

An Exploration Of Linear Classifiers For Unsupervised Spiking Neural Networks With Event-Driven Data, Wesley Chavez

Dissertations and Theses

Object recognition in video has seen giant strides in accuracy improvements in the last few years, a testament to the computational capacity of deep convolutional neural networks. However, this computational capacity of software-based neural networks coincides with high power consumption compared to that of some spiking neural networks (SNNs), up to 300,000 times more energy per synaptic event in IBM's TrueNorth chip, for example. SNNs are also well-suited to exploit the precise timing of event-driven image sensors, which transmit asynchronous "events" only when the luminance of a pixel changes above or below a threshold value. The combination of event-based imagers …

Theory, Operation, And Application Of Neural Networks, John Andersen

University Honors Theses

This paper examines the history and current state of machine learning. It examines neural networks, theory behind neural networks, how they are implemented, and how they are used. The systems and networks examined have up to three modes of learning. Theory behind machine learning is broken up into three approaches; rule-based, Bayesian, and neural networks. Operation of machine learning algorithms has been enabled by several prevalent libraries in the open source community, as well as various hardware technologies. Due to this surge in resources application developers have been able to apply machine learning in novel ways. An application of machine …