STUDENT RESEARCH SHOWCASE 2020 PRESENTATIONS

Math and Computer Science

Submit

Computer Efficiency and Delta Queuing
Ramsey Alsheikh, American Heritage School Boca Delray

Improving computer speed is of vital importance to the development of several new technologies, including autonomous vehicles and surgical robotics. Small delays and lag-time can spell disaster in real time applications such as these. This project proposes one way to improve computer speed through Delta Queuing, an algorithm to enhance the efficiency of process scheduling and sleeping within existing operating systems.

 

Submit

Comparing the Data Distribution of Leading Digits of Ligand Receptors to Benford’s Law to Detect Fraudulent Data
Vardhan Avaradi, American Heritage School Plantation

My presentation details the background behind my project, what it focuses on, my experimental matrix, my results, and a discussion for conclusion and applications of this research project.

Submit

Row,  Row, Row Your Boat
Audrey Biller, East Chapel Hill High School

In this presentation, I analyze world indoor junior rowing championship data. I present the distribution of the 2,000-meter erg times for each junior rowing category and estimate the erg-time thresholds that are necessary for placement as a function of age, gender, and weight class. The junior rowers could use these findings to assess their current performance and understand where they should be by age 18 for top placement in their categories to get farther in the recruitment process. I further develop a predictive model to allow any junior rower to identify her/his future erg-time based on current performance, age, gender and weight class. The data analysis shows that the key feature that determines the erg time of a rower is first gender and then weight class. The analysis also shows that gender and weight class disadvantages may be overcome, to a certain extent, through efficient training programs. The mean erg time decreases with age, independent of gender and weight class. Especially female light-weight rowers should not be discouraged by the smaller early improvements; they typically see larger improvements in their junior and senior years. The most important thing to do for a rower in such a situation is to persevere and believe that hard work will eventually pay off. The presentation concludes with inspiring stories of selected world indoor junior rowers who have succeeded in doing that and significantly improved their erg times over four consecutive years.

Submit

User Preferences in Models Predicting Child Support Payment Delinquency
Carolina Bolnykh, Homeschool

The purpose of this project is to compare the accuracies of predictive models that include social workers' feature preferences on child support delinquency cases and predictive models that do not account for such preferences.

Submit

2-adic Numbers and Quadratic Equations
William Boultinghouse, Kentucky Wesleyan College

This presentation will highlight the key ideas of the authors' research diving into 2-adic valuations and their application to sequences.

Submit

The Efficiency to Which Evolutionary Algorithms Can Decrypt Chaotically Encrypted Images
Tarik Campbell, American Heritage School Boca Delray

My presentation focuses on the application of chaotic algorithms in the field of image encryption and the application of evolutionary algorithms in the field image decryption.

Submit

Prediction of High School Dropout Rates Using Machine Learning
Ira Chaturvedi, University High School

This project uses machine learning to predicted the number of high school students likely to dropout at a school given race, gender, and location.  The final product is in the form of a website where one can put upload information for a certain category or upload a full csv file form.

Submit

PanCan Diagnosed: Developing an Algorithm for the Accurate and Affordable Early Diagnosis of Pancreatic Cancer via Machine Learning and Bioinformatics
Siya Goel, West Lafayette Junior Senior High School

To address the problems of late diagnosis and misdiagnosis of pancreatic cancer (PC), an accessible, accurate, and affordable diagnostic tool was developed by analyzing 19 genes using machine learning and bioinformatics. PanCan diagnosis is the first clinical test that achieves a 92% diagnostic accuracy, potentially increasing PC survival rate from 8.9% to 39.5%.

Submit

Applied Artificial Intelligence: Differentiating Between Glaucomatous, High- and Low-Risk Suspect, and Normal Eyes with Neural Networks
Jay Gopal, North Broward Preparatory School

Glaucoma, an eye disorder with intraocular pressure-associated optic neuropathy, is the leading cause of worldwide irreversible blindness, but it is estimated that half of the people with the disease believe they have healthy eyes! Among the various tools that can increase awareness for glaucoma is artificial intelligence (AI). My research developed and evaluated a convolutional neural network (CNN), a tool often used in deep learning studies to classify images, to distinguish between normal, low-risk suspect, high-risk suspect, and glaucomatous eyes with high accuracy. The long-term vision is to develop an app that people can use at home to instantly determine whether they have glaucoma.

Submit

Peripheral Image Customization using Machine Learning
Ruth Sandra Harry, Christ University

Customization of images involves changing various aspects of the images to optimize some viewer metrics such as view time, click through rate, number of interactions, and so on. However image customization is increasingly expensive since it requires some form of intelligent intervention to choose what aspects of images must be modified and for which viewers. In the age of machine learning and artificial intelligence however, such problems are routine and scalability is easily achieved. In this research study we provide one image customization scheme and study its impact.

Submit

A Data-Driven Assessment of the Link Between Abelson Helper Integration Site-1 (Ahi1) Gene Expression and Stress Sensitivity in Adolescents
Zachary Hine, Bronx High School of Science

I will present how I engineered an XGBoost Machine Learning model and trained it with data from the Adolescent Brain and Cognitive Development study to predict an adolescent's sensitivity level to stress-inducing stimuli.

Submit

AI-Driven Strawberry Harvesting Robot using Custom Farm Survey Model based on SSD with Inception v2
Aditya Indla, Bellarmine College Preparatory School

Strawberry farming is both labor-intensive and time-sensitive. As a result, farmers employ fixed labor to pick fruit every other day, regardless of the number fruit ready to harvest, a massively inefficient process. I thought that by using Artificial Intelligence, Computer Vision and Robotics labor costs could be lowered and farming made more efficient. A custom image recognition model can provide more information on the farm’s status and allow farmers to plan labor needs better. A harvester robot can automate the picking process and reduce the costs further. My solution consists of two parts: A Custom Object Detection Model based on the Single Shot Detector (SSD) Inception V2 Convolutional Neural Network (CNN) Image Recognition Model to identify and classify strawberries based on ripeness; A robot that utilizes this model to identify and pick ripe strawberries.

Submit

Improving Genomic Analysis by Assessing the Causes of Reference Bias in Genome Assembly
Sheila Iyer, Thomas Jefferson High School for Science and Technology

With the advent of next-gen sequencing, DNA sequencing involves massively parallel sequencing and constructing a genome by assembling 50-400 base-pair fragments based on overlaps in sequence. In the read-alignment phase of genome assembly, there are many sources of bias that can affect alignment results. Reference bias is caused by the inherent bias of an aligner to the reference genome sequence. There is a need to understand how reference bias arises and if there are different methods that can be integrated into the read alignment pipeline to reduce allelic bias. By doing so, downstream genetic analysis of hypervariable regions can be improved and result in more accurate disease modeling. In this paper, we use real and simulated reads to calculate reference bias statistics, graphically represent reference bias, and show how reference bias can be reduced. Our results show that aligning to the major allele reference genome or using a reference flow method whereby genetic variants are incorporated stochastically from different population groups can reduce reference bias by 31.4% and 66.3%, respectively. Furthermore, we compare the performance of the linear reference genome aligner Bowtie2 and the graph genome aligner Hisat2, which yield very similar reference bias of 50.48% and 50.27%, respectively.

Submit

ASL Fingerspelling Classification
Moira Katherine Minielly, Pine Crest School

There is a communication barrier between deaf and hearing people, but recently deep learning tools have been utilized to improve American Sign Language (ASL) translation. My presentation focuses on utilizing different image transformation techniques, such as background removal and hand identification, to improve ASL fingerspelling image classification accuracy. This image transformation method can then be applied to real-time ASL translation in the future, and increase the accuracy of real-time ASL translation.

Submit

Automated Recyclable Detection in Industrial Processing Facilities Using Deep Learning
William Kirschner, Pine Crest School 

Given the rising production of solid waste, it is becoming paramount to find ways to reuse existing products. Recycling has long been the solution but, recently, has become increasingly difficult. To address this issue, my presentation explores the application of artificial intelligence and deep learning in terms of automating the sorting process.

Submit

The Application of Compressed Sensing in Neural Cryptography
Alexander Mark, American Heritage School Boca Delray

Neural cryptography is a relatively new field in which adversarial neural networks (ANN) are used to develop encryption systems. Compressed sensing is the use of mathematical transformations on datasets, particularly sparse datasets, to condense the amount of data present without losing any information. The purpose of this experiment was to evaluate the effect of compressed sensing on an adversarial neural network’s ability to encrypt an image in a secure fashion.

Submit

Examining an Iterative Development of Reward Functions to Generate Autonomous Driving Models using Reinforcement Learning
Sashrika Pandey, Irvington High School

This study aimed to examine the development of reinforcement learning models for autonomous navigation using a technique called pre-training, which clones successful models and introduces new reward functions to optimize a virtual car's speed and accuracy over a track. The goals of this study were to optimize a combination of parameters to complete a lap around a track quickly and accurately and to determine the effectiveness of pre-training in contrast to newly initialized models.

Submit

Progress Toward Reducing the Error Rates of Quantum Computers
Dylan Parsons, American Heritage School Boca Delray

This presentation gives an overview of quantum computing and details the lowering of quantum "noise" within qubits (quantum bits) since 2017.

Submit

Generation of Computed Tomography Scans and Segmentation Masks Using Generative Adversarial Networks
Tejas Prabhune, Evergreen Valley High School

Researchers are attempting to benefit doctors by creating AI solutions to various medical problems. However, machine learning requires large amounts of public data that is often unavailable for these researchers. In this study, a framework is proposed to synthesize CT scans and their segmentation masks using a pair of Generative Adversarial Networks (GANs) as an effort to tackle this data scarcity.

Submit

Using Conditional Generative Adversarial Networks to Reduce the Effects of Latency in Robotic Telesurgery
Neil Sachdeva, Pine Crest School

Robotic telesurgery is an upcoming technology that allows for remote control of surgical robots. The problem with this is that latency and lag make remote operations very unsafe for patients as any network delay could result in an accident that injures a patient. My research developed an Artificial Intelligence system that tracks robotic arms in a surgical setting, which the base for an automated reflex system that would prevent any injuries during operation and make robotic telesurgery possible.

Submit

Using Neural Network to Distinguish Between Artificial and Real Leaves
Ritvik Teegavarapu, American Heritage School Boca Delray

Artificial neural networks (ANNs) belong to a class of machine learning approaches under AI which can be trained to recognize patterns and approximate nonlinear functions provided data is available for ANNs to learn. The ability of the ANNs in differentiating an artificial object from a real object is one of the active areas of research.

Submit

Identifying Optimal Panobinostat Treatment Regimens Utilizing Reverse-Engineered Concentration-Time Curves
Jonathan Williams, Pine Crest School

My presentation introduces my mathematical oncology research conducted at the Moffitt Cancer Center.

Submit

Increasing Disaster Relief Efficiency through Machine Learning Algorithms
Sophia Zheng, American Heritage School Boca Delray

The purpose of this experiment was to utilize machine learning to analyze a dataset of migration patterns to determine places where humanitarian relief is most needed after a natural disaster, thus increasing the efficiency of aid distribution. The hypothesis was that if machine learning is implemented to analyze a dataset of population patterns, then it can predict migration patterns after disasters with a higher accuracy than the regular human researcher can and it can increase the efficiency of aid distribution during disaster crises. A dataset from Kaggle titled “Human Mobility During Natural Disasters” was analyzed using Weka software. The dataset consisted of the geographic coordinates of tweets during Typhoon Wipha. Three models were built by leveraging principal component analysis (PCA): K-means, Density based clustering, and Farthest First. As seen by the mappings, there were distinct hotspots of certain coordinates so these models could be used to analyze future test inputs, so the hypothesis was supported in that a machine learning model could be built on the available data to test further data. By creating various machine learning models based on PCA, the researcher can help identify where hotspots of aid would be most necessary after or during a natural disaster which increases the efficiency of humanitarian aid.

Submit