STUDENT RESEARCH SHOWCASE 2019 PRESENTATIONS

Math and Computer Science

The Formulas and Patterns Behind Right Integral Triangle Side Lengths
Spencer Bauman, Pine Crest School

My presentation explores my research regarding right integral triangle side lengths. I created a proof for the formulas I found in past research articles about the same topic, and I created functions that contain all the Pythagorean Triples at its intersection points.

Deep Learning Control System Using an Air-Gapped Optical Input Stream
Rodrigo Castellon, Pine Crest School

The problem of minimizing risk in industrial plant settings is a research interest. Long hours in industrial control rooms induce fatigue, drowsiness, and ease of distraction, posing a risk of failure or catastrophe. Considering the critical importance industrial plants for the modern-day economy, the minimization of risk in such settings should be prioritized. Recently, deep learning has proven successful on a host of applications, including online recommendation systems, online language translation, image recognition, robotics, and medical diagnosis. This research aimed to develop a deep learning control system that learned to attain human performance on a task (a 1990's Microsoft DOS racing game called Car & Driver) with only optical input, serving as a preliminary prototype for controlling a power plant from only optical feed of dials and gauges. A supervised learning approach was taken; a convolutional neural network (CNN) was trained on 6 hours of pre-recorded human expert play data to return the optimal keystroke given a still frame image of the task. When evaluated offline on a test dataset of still image frames of the same task, the CNN achieved 91.5% accuracy. To evaluate the CNN online on the task itself, a novel hardware interface consisting of two connected Arduino boards was developed to facilitate CNN-task interaction. When all components were incorporated into a cohesive control system, the CNN achieved near-human performance on the in-game metrics “Average Speed,” “Top Speed,” and “Lap Time,” demonstrating the system’s capacity to achieve human-performance when applied to an actual power plant scenario.

Melanoma Skin Cancer Identification Using Deep Learning
William Kirschner, Pine Crest School

Using smartphone/tablet attachable dermatoscope to train a Convolutional Neural Network to accurately identify melanoma and benign skin lesions.

The Development of a Data Pipeline to Improve Cancer Diagnosis Using Convolutional Neural Networks
Juliette Koval, Pine Crest School

My presentation discusses a machine learning algorithm I programmed that classifies any skin lesion (mole) as cancerous or benign in order to improve the existing diagnoses, as the process is subjective and creates many inaccuracies. The algorithm correctly diagnoses 99.96% of cancerous moles as cancerous, and is thus significantly more accurate than an average general practitioner (correctly diagnose 37.5% of cancerous moles).

The Effect of Random Noise and Data Abundance on the Accuracy of a Polynomial Approximating Neural Network
Alexander Mark, American Heritage School Boca/Delray

This presentation displays the results and analysis of results of an experiment conducted to evaluate the effects of different types of data on the accuracy of an artificial intelligence model.

Development of a Deep Structured Learning Model to Predict the Functional Impacts of Unknown Cancerous Mutations
Naven Parthasarathy, American Heritage School Boca/Delray

Deep learning, classified as a sub-field of machine learning, utilizes artificial neural networks that enable computers to learn from observational data by using layers of neuron-like nodes that mimic how human brains analyze information. This project took on the problem of taking on extremely large, yet annotated datasets and enabling the computer to extract raw features and construct a productive tool based on patterns that are buried deep within the data. The data was first queried from the database of the International Cancer Genome Consortium which has identified almost eighty million simple somatic mutations and contains information from change in DNA and type of mutation to clinical evidence and exhibited phenotypic effects—as well as hundreds of potential variations and transcripts, each thousands of nucleotide bases long. After filtering through the highly organized yet partly superficial data, sequestered into a SQLite database, it was reorganized into a series of mutation transcripts connected to a sequence of strings and integers that represented the type of mutation, consequences, functional impact, origin, as well as clinical significance. The model will train on this easier-to-process data and learn features of the data which it will use for predictive analysis. The ultimate goal would to be to provide an unclassified, and extremely large, segment of DNA and have the machine predict features about what type of mutation it is and characteristics about this mutation including, for example, potential loss-of-function that may arise in afflicted patients to better treat these mutations which previously have been silent killers.

Developing a Histopathological Image-Based Tool for Accurate Prognosis of Pancreatic Adenocarcinoma by Implementing Deep Structured Learning
Anish Ravichandran, American Heritage School Boca/Delray

A PowerPoint and video describing how whole slide images and a deep learning network were used to create an accurate prognosis for pancreatic cancer patients.

Counting Past Infinity: A Study on Cantor's Continuum Hypothesis
William Rhodes, American Heritage School Boca/Delray

This experiment sought to answer the age old Continuum Hypothesis. While unsuccessful, it provided insight for coming to a conclusion in the future.

Prime Time - Primality Testing With Proof of Correctness: Maurer’s Algorithm vs. Miller-Rabin
Rose Rothschild, Pine Crest School

My research compares different prime number algorithms to see which one is more efficient and secure.

Impact of Network Disruption on Optimal Pathlength and Transport Efficiency
Supriya Roy, Mayo High School

Resource distribution networks are used everywhere to transport people, energy, resources, and information. This presentation describes a mathematical model of network design inspired by a slime mold that is efficient yet robust. Scientists and engineers can use this mathematical model to design efficient and sustainable transport networks that aren’t vulnerable to disruption.

Using Generative Machine Learning Models to Establish a Relation Between Deaf and Normal Speech Reflective of Hearing Deficiencies
JayaSAI Somasundaram, American Heritage School Plantation

In America nearly 10 million people are hard of hearing and of these 10 million people about 1 million are functionally deaf (can’t hear at all). Current methods of testing done by audiologist is limited and causes a bottleneck for the usefulness of Cochlear Implants and other Hearing Aids(B. Banerjee, 2016). The goal of the project is to solve for this issue and give audiologist a comprehensive system of measurement for deficiencies in hearing. The system should be easy to use and quick. The project does this by training auto-encoding machine learning models and plotting the intermediate representations of the input. These models are trained individually on the deaf speech and regular speech. The intermediate representation are plotted and analytically compared to find the deficiencies. The results of this project showed that a special type of Recurrent Neural Network could be modified and made to work for this application. Different tests such as comparisons of the same clips of audio were compared to verify that the outputs were reliable and accurate. Tests on patients are also trying to be conducted. This project proves that parameters of models if manipulated correctly can be intermediate representations and that speech when manipulated can be representative of hearing. The application includes a completely revolutionized system for isolating hearing deficiencies as well as a model developed to tune cochlear aid using the output of this project.

Mental Disorder Diagnosis Based on Single-photon Emission Computed Tomography using Deep Neural Networks
Oleksii Volkovskyi, American Heritage School Plantation

I created a machine learning algorithm that diagnoses the mental disorder of a patient based on the single-photon emission computed tomography scan of their brain. It looks at the activity levels throughout different regions of the brain and uses them as inputs into a dense neural network to classify the general mental disorder and classify between anxiety and ADHD.

STUDENT RESEARCH SHOWCASE 2019 PRESENTATIONS

Math and Computer Science

About

Programs

Chapters

Meetings & Events

Members

News

Publications